" + "
".join(extra_parts) + "
Last updated 2026-03-02
- - -Last updated 2026-03-02
- - -This archived example focused on building an agent with `smolagents` on top of a Hugging Face deployment in Microsoft Foundry. - -## What it covered - -- Deploying a coding-oriented LLM from the Hugging Face collection -- Invoking the model from Python -- Wiring the deployment into a `smolagents` workflow -- Comparing Hub, Azure ML, and Foundry views of the same model - -## Original assets - -- [Notebook source](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/build-agents-with-smolagents/azure-notebook.ipynb) -- [Example directory](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/build-agents-with-smolagents) - -## Current docs to read instead - -- [Build agents with Foundry deployments](../../../agents/introduction) -- [Define custom tools](../../../agents/define-custom-tools) -- [Connect MCP servers](../../../agents/connect-mcp) \ No newline at end of file diff --git a/docs/source/deprecated/foundry/examples/deploy-large-language-models.mdx b/docs/source/deprecated/foundry/examples/deploy-large-language-models.mdx deleted file mode 100644 index 9e8b004..0000000 --- a/docs/source/deprecated/foundry/examples/deploy-large-language-models.mdx +++ /dev/null @@ -1,27 +0,0 @@ -# Archived Example: Deploy Large Language Models on Microsoft Foundry - -Last updated 2026-03-02
- - -Last updated 2026-03-02
- - -This archived example showed a more notebook-heavy path for deploying a text-generation model and then calling it from several clients. - -## What it covered - -- Programmatic deployment with the Azure ML SDK -- OpenAI-compatible inference calls -- A local Gradio chat client -- The relationship between the Hugging Face model ID and the Azure model URI - -## Original assets - -- [Notebook source](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-large-language-models/azure-notebook.ipynb) -- [Example directory](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-large-language-models) - -## Current docs to read instead - -- [Deploy your first model](../../../get-started/deploy-first-model) -- [Call your deployment](../../../get-started/call-your-deployment) -- [Third-party integrations](../../../guides/third-party-integrations) \ No newline at end of file diff --git a/docs/source/deprecated/foundry/examples/deploy-nvidia-parakeet-asr.mdx b/docs/source/deprecated/foundry/examples/deploy-nvidia-parakeet-asr.mdx deleted file mode 100644 index 6ead456..0000000 --- a/docs/source/deprecated/foundry/examples/deploy-nvidia-parakeet-asr.mdx +++ /dev/null @@ -1,26 +0,0 @@ -# Archived Example: Deploy NVIDIA Parakeet for ASR on Microsoft Foundry - -Last updated 2026-03-02
- - -Last updated 2026-03-02
- - -This archived example focused on automatic speech recognition using NVIDIA Parakeet. - -## What it covered - -- Deploying an ASR model from the Hugging Face collection -- Sending audio transcription requests -- Comparing different client approaches, including cURL and Python -- Working with a speech-focused deployment instead of a text chat model - -## Original assets - -- [Notebook source](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-nvidia-parakeet-asr/azure-notebook.ipynb) -- [Example directory](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-nvidia-parakeet-asr) - -## Current docs to read instead - -- [Deploy your first model](../../../get-started/deploy-first-model) -- [Supported tasks](../../../reference/supported-tasks) \ No newline at end of file diff --git a/docs/source/deprecated/foundry/examples/deploy-smollm3.mdx b/docs/source/deprecated/foundry/examples/deploy-smollm3.mdx deleted file mode 100644 index de6ce17..0000000 --- a/docs/source/deprecated/foundry/examples/deploy-smollm3.mdx +++ /dev/null @@ -1,26 +0,0 @@ -# Archived Example: Deploy SmolLM3 on Microsoft Foundry - -Last updated 2026-03-02
- - -Last updated 2026-03-02
- - -This archived example focused on deploying `HuggingFaceTB/SmolLM3-3B` and calling it through the OpenAI-compatible route. - -## What it covered - -- Building the Azure model URI from the Hugging Face model ID -- Creating a managed endpoint and deployment -- Running text-generation requests -- Reasoning-oriented and long-context positioning for a small model - -## Original assets - -- [Notebook source](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-smollm3/azure-notebook.ipynb) -- [Example directory](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-smollm3) - -## Current docs to read instead - -- [Deploy your first model](../../../get-started/deploy-first-model) -- [Thinking levels](../../../guides/thinking-levels) \ No newline at end of file diff --git a/docs/source/deprecated/foundry/examples/deploy-vision-language-models.mdx b/docs/source/deprecated/foundry/examples/deploy-vision-language-models.mdx deleted file mode 100644 index 053525d..0000000 --- a/docs/source/deprecated/foundry/examples/deploy-vision-language-models.mdx +++ /dev/null @@ -1,26 +0,0 @@ -# Archived Example: Deploy Vision Language Models on Microsoft Foundry - -Last updated 2026-03-02
- - -Last updated 2026-03-02
- - -This archived example focused on multimodal deployments, using a vision-language model that accepts both images and text. - -## What it covered - -- Deploying a VLM from the Hugging Face collection -- Sending multimodal requests -- Using the deployment through both SDK and notebook flows -- Handling image-aware prompts in a Foundry deployment - -## Original assets - -- [Notebook source](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-vision-language-models/azure-notebook.ipynb) -- [Example directory](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/foundry/deploy-vision-language-models) - -## Current docs to read instead - -- [Deploy your first model](../../../get-started/deploy-first-model) -- [Supported tasks](../../../reference/supported-tasks) \ No newline at end of file diff --git a/docs/source/deprecated/foundry/hardware.mdx b/docs/source/deprecated/foundry/hardware.mdx deleted file mode 100644 index 0d18581..0000000 --- a/docs/source/deprecated/foundry/hardware.mdx +++ /dev/null @@ -1,5 +0,0 @@ -# Archived: Supported Hardware - -This page preserves the older hardware table references for GPU and CPU instance families used in Foundry-related deployment flows. - -The maintained version is [Supported hardware](../../reference/supported-hardware). diff --git a/docs/source/deprecated/foundry/introduction.mdx b/docs/source/deprecated/foundry/introduction.mdx deleted file mode 100644 index 4048557..0000000 --- a/docs/source/deprecated/foundry/introduction.mdx +++ /dev/null @@ -1,5 +0,0 @@ -# Archived: Hugging Face on Microsoft Foundry - -This page preserves the former service-introduction page that described the broader Hugging Face and Microsoft collaboration, the expansion through Build and Ignite announcements, and the catalog-scale messaging used in the earlier docs. - -That material has been replaced in the main docs by a shorter, task-oriented [Overview](../../index). diff --git a/docs/source/deprecated/foundry/models.mdx b/docs/source/deprecated/foundry/models.mdx deleted file mode 100644 index 6e3721e..0000000 --- a/docs/source/deprecated/foundry/models.mdx +++ /dev/null @@ -1,9 +0,0 @@ -# Archived: Supported Models - -This page preserves the earlier explanation of how to check whether a Hugging Face model is available in the Foundry catalog, including: - -- Checking the **Deploy on Microsoft Foundry** action on the Hub -- Searching the Hugging Face collection in Foundry -- Using the old Azure lookup endpoint mentioned in the earlier docs - -The maintained version is [Supported models](../../reference/supported-models). diff --git a/docs/source/deprecated/foundry/tasks.mdx b/docs/source/deprecated/foundry/tasks.mdx deleted file mode 100644 index 2ac9d59..0000000 --- a/docs/source/deprecated/foundry/tasks.mdx +++ /dev/null @@ -1,5 +0,0 @@ -# Archived: Supported Tasks - -This page preserves the older task list wording that grouped supported workloads across text, multimodal, audio, and vision. - -The maintained version is [Supported tasks](../../reference/supported-tasks). diff --git a/docs/source/deprecated/guides/access-gated-models.mdx b/docs/source/deprecated/guides/access-gated-models.mdx deleted file mode 100644 index a964851..0000000 --- a/docs/source/deprecated/guides/access-gated-models.mdx +++ /dev/null @@ -1,13 +0,0 @@ -# Archived: Deploy Hugging Face gated models on Microsoft Foundry - -This page preserves the former gated-model deployment flow. - -## Legacy flow summary - -1. Request access to the gated repository on the Hugging Face Hub -2. Create a Hugging Face token with gated-repository read permissions -3. Store that token in an Azure Machine Learning connection -4. Ensure the endpoint can access the secret store -5. Deploy the model using the Azure ML SDK - -The maintained version of this guidance is [Gated models](../../guides/gated-models). diff --git a/docs/source/deprecated/guides/configure-azure-ml-microsoft-foundry.mdx b/docs/source/deprecated/guides/configure-azure-ml-microsoft-foundry.mdx deleted file mode 100644 index e91c641..0000000 --- a/docs/source/deprecated/guides/configure-azure-ml-microsoft-foundry.mdx +++ /dev/null @@ -1,17 +0,0 @@ -# Archived: Configure Azure Machine Learning and Microsoft Foundry - -This page preserves the former setup flow that walked through: - -1. Installing and authenticating the Azure CLI -2. Installing the `az ml` extension -3. Creating a resource group -4. Creating an Azure AI Foundry hub -5. Creating a project attached to that hub - -That flow assumed the older Azure AI Foundry naming and hub/project model. The maintained version is [Configure Microsoft Foundry](../../get-started/configure-foundry). - -## Legacy notes - -- Older guides and notebooks may refer to Azure AI Foundry instead of Microsoft Foundry -- Older commands may reference hub-based projects explicitly -- Some screenshots and resource names in notebooks still follow that older wording diff --git a/docs/source/deprecated/guides/introduction.mdx b/docs/source/deprecated/guides/introduction.mdx deleted file mode 100644 index fbd7cb1..0000000 --- a/docs/source/deprecated/guides/introduction.mdx +++ /dev/null @@ -1,17 +0,0 @@ -# Archived: Guides - -This page used to act as the guides index in the former documentation structure. - -The old guide set centered on: - -- One-click deployments from the Hugging Face Hub -- Deploying gated models -- Requesting model additions -- Configuring Azure Machine Learning and Microsoft Foundry -- Managed compute - -The maintained equivalents now live under: - -- [Get started](../../get-started/introduction) -- [Guides](../../guides/managed-compute) -- [Agents](../../agents/introduction) diff --git a/docs/source/deprecated/guides/managed-compute-foundry.mdx b/docs/source/deprecated/guides/managed-compute-foundry.mdx deleted file mode 100644 index da79f67..0000000 --- a/docs/source/deprecated/guides/managed-compute-foundry.mdx +++ /dev/null @@ -1,12 +0,0 @@ -# Archived: Deploy Hugging Face Models in Foundry with Managed Compute - -This archived page keeps the older, more detailed managed-compute framing that predated the simplified guide set. - -The former guide emphasized: - -- Managed compute as a dedicated-capacity deployment option -- Accelerator-family-specific quota requirements -- Template-driven deployment choices -- Validation through the Foundry playground before moving to code - -For the current streamlined version, see [Managed compute](../../guides/managed-compute). diff --git a/docs/source/deprecated/guides/one-click-deployment-foundry.mdx b/docs/source/deprecated/guides/one-click-deployment-foundry.mdx deleted file mode 100644 index 7f79db7..0000000 --- a/docs/source/deprecated/guides/one-click-deployment-foundry.mdx +++ /dev/null @@ -1,17 +0,0 @@ -# Archived: One-click deployments from the Hugging Face Hub on Microsoft Foundry - -This page preserves the former one-click deployment entry point. - -## Legacy flow - -1. Browse models on the Hugging Face Hub -2. Use the **Deploy** button on a supported model card -3. Choose **Deploy on Microsoft Foundry** -4. Review the model card in Foundry -5. Create the deployment and test the endpoint - -The current maintained equivalent is [Deploy your first model](../../get-started/deploy-first-model). - -## Why keep this archived page? - -Older tutorials, screenshots, and issues often refer to the Hub-first flow rather than the current docs navigation. diff --git a/docs/source/deprecated/guides/request-model-addition.mdx b/docs/source/deprecated/guides/request-model-addition.mdx deleted file mode 100644 index cbf61ee..0000000 --- a/docs/source/deprecated/guides/request-model-addition.mdx +++ /dev/null @@ -1,13 +0,0 @@ -# Archived: Request a model addition in the Hugging Face collection on Microsoft Foundry - -This page preserves the earlier model-addition guidance from the old docs structure. - -The previous checklist focused on: - -- `Transformers`, `Diffusers`, or `Sentence-Transformers` compatibility -- A supported task taxonomy -- Runtime alignment for TGI, vLLM, SGLang, or TEI when relevant -- Public visibility on the Hub -- Secure artifact formats and security scanning - -The maintained version is [Request model addition](../../guides/request-model-addition). diff --git a/docs/source/deprecated/machine-learning/examples/tool-calling-agent-fine-tuning.mdx b/docs/source/deprecated/machine-learning/examples/tool-calling-agent-fine-tuning.mdx deleted file mode 100644 index b74f669..0000000 --- a/docs/source/deprecated/machine-learning/examples/tool-calling-agent-fine-tuning.mdx +++ /dev/null @@ -1,22 +0,0 @@ -# Archived Example: Tool-calling Agent Fine-tuning on Azure Machine Learning - -Last updated 2026-03-02
- - -Last updated 2026-03-02
- - -This archived example comes from the earlier Azure Machine Learning example set rather than the current Foundry-first docs path. - -## What it covered - -- Fine-tuning an agent-oriented model -- Preparing training data for tool-calling behavior -- Running an Azure Machine Learning job -- Inspecting experiment and log outputs - -## Original assets - -- [Example directory](https://github.com/huggingface/Microsoft-Azure/tree/main/examples/machine-learning/tool-calling-agent-fine-tuning) - -This material remains archived because the active docs are now centered on Foundry deployment and integration workflows instead of Azure ML training examples. \ No newline at end of file diff --git a/docs/source/deprecated/overview.mdx b/docs/source/deprecated/overview.mdx deleted file mode 100644 index a266581..0000000 --- a/docs/source/deprecated/overview.mdx +++ /dev/null @@ -1,11 +0,0 @@ -# Deprecated / Archived - -This section preserves the former documentation layout and legacy examples that existed before the docs were reorganized around the simpler Foundry-focused structure. - -Use this section when you need to: - -- Follow an old link that no longer belongs in the main learning path -- Recover older wording, screenshots, or step-by-step examples -- Compare the previous docs organization with the current one - -The archived pages are kept for continuity, not as the recommended entry point for new users. For the current docs, start from the main [Overview](../index). diff --git a/docs/source/deprecated/resources.mdx b/docs/source/deprecated/resources.mdx deleted file mode 100644 index 6dcf149..0000000 --- a/docs/source/deprecated/resources.mdx +++ /dev/null @@ -1,14 +0,0 @@ -# Archived: Other Resources - -This page keeps the older external-reading collection that used to live near the top of the docs. - -## Legacy links - -- [Hugging Face on Azure](https://azure.microsoft.com/en-us/solutions/hugging-face-on-azure) -- [Microsoft and Hugging Face expand collaboration](https://huggingface.co/blog/azure-ai-foundry) -- [Microsoft and Hugging Face expand collaboration to accelerate Open-Source AI Innovation on Azure AI Foundry](https://devblogs.microsoft.com/foundry/microsoft-and-hugging-face-expand-partnership-to-accelerate-open-source-ai-innovation-on-azure-ai-foundry/) -- [From cloud to developers: Hugging Face and Microsoft deepen collaboration](https://huggingface.co/blog/microsoft-collaboration) -- [Microsoft and Hugging Face deepen generative AI partnership](https://techcommunity.microsoft.com/blog/aiplatformblog/microsoft-and-hugging-face-deepen-generative-ai-partnership/4144565) -- [Hugging Face collaborates with Microsoft to launch Hugging Face Model Catalog on Azure](https://huggingface.co/blog/hugging-face-endpoints-on-azure) - -For the maintained external-links page, see [Resources](../resources). diff --git a/docs/source/deprecated/security.mdx b/docs/source/deprecated/security.mdx deleted file mode 100644 index 4d552b4..0000000 --- a/docs/source/deprecated/security.mdx +++ /dev/null @@ -1,14 +0,0 @@ -# Archived: Security & Compliance - -This page preserves the earlier security-focused explanation from the old docs layout. - -The legacy guidance highlighted the following model eligibility requirements: - -- Public availability on the Hugging Face Hub, including gated public repositories -- Safe weight formats such as `safetensors` -- No unreviewed `trust_remote_code` or unsupported custom execution code -- Security scanning on Hub artifacts before catalog inclusion - -It also stressed that enterprise controls such as private networking, auditability, and access control still depend on the surrounding Microsoft Foundry and Azure environment. - -For the maintained version of this material, see [Security and model eligibility](../reference/security-and-eligibility). diff --git a/docs/source/get-started/deploy-open-models-on-microsoft-foundry-with-python.mdx b/docs/source/get-started/deploy-open-models-on-microsoft-foundry-with-python.mdx index 2d6f9c8..1fecd9d 100644 --- a/docs/source/get-started/deploy-open-models-on-microsoft-foundry-with-python.mdx +++ b/docs/source/get-started/deploy-open-models-on-microsoft-foundry-with-python.mdx @@ -1 +1,98 @@ -# Deploy open-models from Microsoft Foundry with Python +# Deploy open-models on Microsoft Foundry with Python + +Deploy open models on dedicated hardware with Microsoft Foundry Managed Compute. This example deploys [`zai-org/GLM-5.2-FP8`](https://huggingface.co/zai-org/GLM-5.2-FP8) on 8 AMD Instinct MI300X GPUs. + +## Prerequisites + +- An Azure subscription with Managed Compute preview access +- A Microsoft Foundry account +- `Cognitive Services Contributor` access to the account +- Managed Compute quota for MI300X +- Python 3.10 or later and the Azure CLI + +Sign in and install the dependencies: + +```bash +az login +uv pip install azure-identity "azure-mgmt-cognitiveservices>=15.0.0b2" openai +``` + +## Deploy the model + +Replace the placeholders, then run: + +```python +from azure.identity import DefaultAzureCredential +from azure.mgmt.cognitiveservices import CognitiveServicesManagementClient + +SUBSCRIPTION_ID = "
+
+ 
+
+ Codex
+ OpenAI's coding agent for the terminal
+
+
+
+
+ 
+
+ Hermes Agent
+ Nous Research's self-improving agent
+
+
+
+
From efccbaadf8c1503d47bbe6f6e3a32c3ac185a075 Mon Sep 17 00:00:00 2001
From: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
Date: Wed, 19 Aug 2026 12:42:46 +0200
Subject: [PATCH 11/17] Fix `docs/source/harness/index.mdx`
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/harness/index.mdx | 3 ---
1 file changed, 3 deletions(-)
diff --git a/docs/source/harness/index.mdx b/docs/source/harness/index.mdx
index a09a746..dd2b217 100644
--- a/docs/source/harness/index.mdx
+++ b/docs/source/harness/index.mdx
@@ -23,7 +23,6 @@ Pick a harness to connect it to your Foundry deployment:
OpenAI's coding agent for the terminal
-
Nous Research's self-improving agent
-
+
+ Codex
+ OpenAI's coding agent for the terminal
+
+
+
+
+
+
+ Hermes Agent
+ Nous Research's self-improving agent
+
+
+
+
+ 
+
+ Claude Code
+ Coming soon
+
+
+
+
+
+ Claude Code
+ Coming soon
+
+
+ 
+
+ Pi
+ Coming soon
+
+
+
+
+ Pi
+ Coming soon
+
+ Coming soon
-
Date: Wed, 19 Aug 2026 12:46:30 +0200
Subject: [PATCH 12/17] Fix `overview` and `harness` index pages
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/_toctree.yml | 6 ++----
.../source/large-language-models/deepseek-v4-flash-0731.mdx | 2 +-
2 files changed, 3 insertions(+), 5 deletions(-)
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
index 7b6be11..0612036 100644
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -1,6 +1,5 @@
- sections:
- - local: index
- title: Overview
+ local: index
title: Overview
- sections:
- local: get-started/quickstart
@@ -15,8 +14,6 @@
title: Qwen3.8 27B
title: Large Language Models
- sections:
- - local: harness/index
- title: Overview
- local: harness/claude-code
title: Claude Code (Coming soon)
- local: harness/codex
@@ -25,4 +22,5 @@
title: Hermes Agent
- local: harness/pi
title: Pi (Coming soon)
+ local: harness/index
title: Harness
diff --git a/docs/source/large-language-models/deepseek-v4-flash-0731.mdx b/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
index ac23e2d..9ccf596 100644
--- a/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
+++ b/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
@@ -125,6 +125,6 @@ Use `low` for quick tasks, `high` for most agentic work, and `max` for the harde
## Bonus: Use it in an agent harness
-Connect the deployment to one of the [agent harness integrations](../agents/introduction#harness-integrations). Confirm that the harness supports Chat Completions and the model's reasoning controls.
+Connect the deployment to a compatible harness, such as [Hermes Agent](../harness/hermes-agent). Confirm that the harness supports Chat Completions and the model's reasoning controls.
For more model and serving details, see the [DeepSeek model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731) and the [vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4-Flash).
From 248691bc7bacbdf02bea01f7174553c5c2d05255 Mon Sep 17 00:00:00 2001
From: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
Date: Wed, 19 Aug 2026 13:18:07 +0200
Subject: [PATCH 13/17] Add more custom HTML (WIP)
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/_toctree.yml | 16 +-
docs/source/advanced/reasoning.mdx | 64 +++
docs/source/catalog/deepseek-ai.mdx | 30 ++
docs/source/catalog/qwen.mdx | 531 +++++++++++++++++++++++++
docs/source/get-started/quickstart.mdx | 109 +++--
docs/source/index.mdx | 77 ++--
6 files changed, 735 insertions(+), 92 deletions(-)
create mode 100644 docs/source/advanced/reasoning.mdx
create mode 100644 docs/source/catalog/deepseek-ai.mdx
create mode 100644 docs/source/catalog/qwen.mdx
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
index 0612036..4a82cd0 100644
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -1,6 +1,13 @@
- sections:
- local: index
+ - local: index
+ title: Overview
title: Overview
+- sections:
+ - local: catalog/qwen
+ title: Qwen
+ - local: catalog/deepseek-ai
+ title: DeepSeek AI
+ title: Catalog
- sections:
- local: get-started/quickstart
title: Quickstart
@@ -14,6 +21,8 @@
title: Qwen3.8 27B
title: Large Language Models
- sections:
+ - local: harness/index
+ title: Overview
- local: harness/claude-code
title: Claude Code (Coming soon)
- local: harness/codex
@@ -22,5 +31,8 @@
title: Hermes Agent
- local: harness/pi
title: Pi (Coming soon)
- local: harness/index
title: Harness
+- sections:
+ - local: advanced/reasoning
+ title: Enable and disable reasoning
+ title: Advanced
diff --git a/docs/source/advanced/reasoning.mdx b/docs/source/advanced/reasoning.mdx
new file mode 100644
index 0000000..27fed92
--- /dev/null
+++ b/docs/source/advanced/reasoning.mdx
@@ -0,0 +1,64 @@
+# Enable and disable reasoning
+
+Many of the models in the collection are hybrid reasoning models: they can answer directly, or think first and then answer. The mode is chosen per request, so a single deployment covers both, and you do not need to redeploy to switch.
+
+Reasoning costs latency and output tokens, so it is worth turning off for classification, extraction, or short chat turns, and on for planning, coding, and agentic work.
+
+## Chat Completions
+
+The switch lives in the model's chat template, and vLLM exposes the template arguments through `extra_body`:
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+ base_url="https://.services.ai.azure.com/openai/v1",
+ api_key="",
+)
+
+response = client.chat.completions.create(
+ model="",
+ messages=[{"role": "user", "content": "Sort these tickets by urgency."}],
+ extra_body={"chat_template_kwargs": {"enable_thinking": False}},
+)
+
+print(response.choices[0].message.content)
+```
+
+The argument name is defined by the model, not by the API, so check the model page before using it:
+
+| Model | Argument | Values |
+|---|---|---|
+| Qwen3 and Qwen3.x | `enable_thinking` | `True`, `False` |
+| DeepSeek V4 Flash 0731 | `thinking`, `reasoning_effort` | `True` / `False`, and `low`, `high`, `max` |
+
+When reasoning is enabled and the deployment runs with a reasoning parser, the thinking trace comes back separately from the answer, in `reasoning_content`:
+
+```python
+print(response.choices[0].message.reasoning_content)
+print(response.choices[0].message.content)
+```
+
+Send the `content` of an assistant turn back in the conversation history, but not its `reasoning_content`.
+
+## Responses API
+
+Models served through the Responses API take a reasoning effort instead of a boolean:
+
+```python
+response = client.responses.create(
+ model="",
+ input="Build a small FastAPI health-check endpoint.",
+ reasoning={"effort": "high"},
+)
+
+print(response.output_text)
+```
+
+
+ Agent harnesses usually set the reasoning effort themselves, so configure it there rather than per request. See [Codex](../harness/codex) or [Hermes Agent](../harness/hermes-agent).
+
+
+## Models without a switch
+
+Some models are reasoning-only, such as [QwQ-32B](../catalog/qwen) and Qwen3-30B-A3B-Thinking-2507, and others never emit a thinking trace, such as the `Instruct` variants. Passing the argument to those models has no effect on the output.
diff --git a/docs/source/catalog/deepseek-ai.mdx b/docs/source/catalog/deepseek-ai.mdx
new file mode 100644
index 0000000..b535106
--- /dev/null
+++ b/docs/source/catalog/deepseek-ai.mdx
@@ -0,0 +1,30 @@
+# DeepSeek AI
+
+[DeepSeek AI](https://huggingface.co/deepseek-ai) builds open mixture-of-experts models for reasoning, coding, and agentic workloads, with long context windows and efficient inference.
+
+Every card links to the latest version of the model in the `azure-huggingface` registry on Microsoft Foundry, and shows its input and output modalities:
+
+
+ Managed compute runs on dedicated GPU capacity and bills for as long as the deployment exists. Delete the deployments that you no longer need.
+
-## Choose a deployment path
+## Deploy a model
-| Foundry portal | Python SDK |
-|---|---|
-| The quickest path for a first deployment. [Start in the portal](#foundry-portal). | Best for automation and repeatable environments. [Start with Python](#python-sdk). |
+Pick the Foundry portal for a first, guided deployment, or Python for a repeatable one. Both create the same managed compute deployment.
-Both paths create the same type of managed compute deployment. Choose one, then continue to [Test the deployment](#test-the-deployment).
-
-## Foundry portal
+
+
1. Open [Microsoft Foundry](https://ai.azure.com) and select your project.
-2. Go to **Discover → Models**. Filter **Collection** by **Hugging Face** and **Deployment options** by **Managed compute**.
-3. Select a model, choose **Deploy**, and enter a deployment name. Keep one instance for your first deployment, then select a compatible deployment template and accelerator.
-4. Select **Deploy** and wait for the deployment state to become `Succeeded`.
+2. Go to **Discover → Models**, and filter **Collection** by **Hugging Face** and **Deployment options** by **Managed compute**.
+3. Select a model and choose **Deploy**.
+4. Enter a deployment name, pick a deployment template and its accelerator, and select **Deploy**.
-Provisioning often takes 10–20 minutes and can take longer when accelerator capacity is limited.
+The deployment is ready when its state becomes `Succeeded`.
-## Python SDK
+
+
-Install the dependencies and sign in to Azure:
+Install the dependencies and sign in:
```bash
pip install azure-identity "azure-mgmt-cognitiveservices>=15.0.0b2" openai
az login
```
-Open the model's deployment wizard in Foundry and copy its **Model ID**, **Deployment template ID**, and **Accelerator type**. Replace the placeholders below, then run the script:
+Copy the **Model ID**, **Deployment template ID**, and **Accelerator type** from the model's deployment wizard in Foundry, then run:
```python
from azure.identity import DefaultAzureCredential
from azure.mgmt.cognitiveservices import CognitiveServicesManagementClient
-SUBSCRIPTION_ID = ""
-RESOURCE_GROUP = ""
-ACCOUNT_NAME = ""
-DEPLOYMENT_NAME = ""
-
-MODEL = ""
-TEMPLATE = ""
-ACCELERATOR_TYPE = ""
+SUBSCRIPTION_ID = ""
+RESOURCE_GROUP = ""
+ACCOUNT_NAME = ""
+DEPLOYMENT_NAME = "qwen--qwen3-8-27b"
-management_client = CognitiveServicesManagementClient(
+client = CognitiveServicesManagementClient(
credential=DefaultAzureCredential(),
subscription_id=SUBSCRIPTION_ID,
)
-deployment = management_client.managed_compute_deployments.begin_create_or_update(
+deployment = client.managed_compute_deployments.begin_create_or_update(
resource_group_name=RESOURCE_GROUP,
account_name=ACCOUNT_NAME,
deployment_name=DEPLOYMENT_NAME,
resource={
"sku": {"name": "GlobalManagedCompute", "capacity": 1},
"properties": {
- "model": MODEL,
- "deploymentTemplate": TEMPLATE,
- "acceleratorType": ACCELERATOR_TYPE,
+ "model": "azureml://registries/azure-huggingface/models/qwen--qwen3.8-27b/versions/2",
+ "deploymentTemplate": "azureml://registries/azure-huggingface/deploymenttemplates/qwen--qwen3-8-27b--256k-nvidia-h100/labels/latest",
+ "acceleratorType": "H100_80GB",
"versionUpgradeOption": "OnceNewDefaultVersionAvailable",
},
},
@@ -77,55 +72,43 @@ deployment = management_client.managed_compute_deployments.begin_create_or_updat
print(deployment.properties.provisioning_state)
```
-The call waits for provisioning to finish. The deployment is ready when it prints `Succeeded`.
-
-## Test the deployment
-
-Open the deployment in Foundry and select **Playground**. Send a simple prompt, such as:
-
-```text
-Explain what this model is best suited for in three short bullet points.
-```
-
-If the model responds, the deployment is ready. Applications must use the **deployment name** in the API's `model` field, not the model name from the catalog.
+The call returns when provisioning finishes, and prints `Succeeded`.
-## Call the model from Python
+
+
-Copy the endpoint and API key from the deployment page, then install the OpenAI client if you did not use the Python deployment path:
+
+ The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
+
-```bash
-pip install openai
-```
+## Call the deployment
-For a model that supports the Responses API:
+Copy the endpoint and key from the deployment page, and use the **deployment name** as the `model`:
```python
from openai import OpenAI
client = OpenAI(
- base_url="https://.services.ai.azure.com/openai/v1",
- api_key="",
+ base_url="https://.services.ai.azure.com/openai/v1",
+ api_key="",
)
-response = client.responses.create(
- model="",
- input="Write a one-sentence hello from Microsoft Foundry.",
+response = client.chat.completions.create(
+ model="",
+ messages=[{"role": "user", "content": "Write a one-sentence hello from Microsoft Foundry."}],
)
-print(response.output_text)
+print(response.choices[0].message.content)
```
-Check the model card or deployment page for its supported API. Some models use Chat Completions instead of Responses.
+Some models are served through the Responses API instead, at `/openai/v1/responses`. Check the model page for the API that it supports.
## Clean up
-Delete deployments that you are not using to release their accelerators and stop billing.
-
-- In Foundry, open the deployments list, select the deployment, and choose **Delete**.
-- If you deployed with Python, use the management client created earlier:
+Delete the deployment to release its accelerators and stop billing, either from its page in Foundry or with:
```python
-management_client.managed_compute_deployments.begin_delete(
+client.managed_compute_deployments.begin_delete(
resource_group_name=RESOURCE_GROUP,
account_name=ACCOUNT_NAME,
deployment_name=DEPLOYMENT_NAME,
@@ -134,6 +117,6 @@ management_client.managed_compute_deployments.begin_delete(
## Next steps
-- Try a model-specific guide for [GLM 5.2 FP8](../large-language-models/glm-5-2-fp8) or [DeepSeek V4 Flash 0731](../large-language-models/deepseek-v4-flash-0731).
-- Connect a Responses-compatible deployment to [Codex](../harness/codex) or another supported harness.
-- See the [Managed compute documentation](https://learn.microsoft.com/azure/foundry/how-to/deploy-models-managed) for quota, monitoring, scaling, and troubleshooting.
+- Browse the [catalog](../catalog/qwen) to pick a model.
+- Turn reasoning on or off per request in [Enable and disable reasoning](../advanced/reasoning).
+- Connect the deployment to a harness such as [Codex](../harness/codex) or [Hermes Agent](../harness/hermes-agent).
diff --git a/docs/source/index.mdx b/docs/source/index.mdx
index 6964bc3..acd2a71 100644
--- a/docs/source/index.mdx
+++ b/docs/source/index.mdx
@@ -2,49 +2,72 @@
-Hugging Face and Microsoft work together to make open models easier to discover and deploy on Microsoft Foundry. The Hugging Face collection in Foundry brings a large catalog of open models into the Foundry experience so teams can deploy them from the UI or from Python without building the full serving stack themselves.
+Hugging Face and Microsoft bring a curated collection of open models to Microsoft Foundry. You deploy a model on managed compute from the Foundry portal or from Python, and call it behind an OpenAI-compatible endpoint, without building or maintaining the serving stack yourself.
-With Hugging Face on Microsoft Foundry, you can:
+## Catalog
-- Discover open models from the Hugging Face catalog directly in Microsoft Foundry.
-- Deploy supported models from the UI or programmatically from Python.
-- Use the resulting deployments behind Foundry-managed endpoints.
-- Build applications, agents, and examples on top of those deployments.
+The collection lives in the `azure-huggingface` registry and shows up in the Foundry model catalog, where you can filter by **Collection: Hugging Face** and **Deployment options: Managed compute**.
-## Microsoft Foundry
+
-The Hugging Face collection is available from the Foundry model catalog, where you can filter models by family, modality, task, and deployment option before starting a deployment.
+Each model is published as a versioned asset in the registry, with one or more deployment templates that pin the serving engine, the context window, and the accelerator it runs on. Browse the collection by family in the [catalog](./catalog/qwen), where every entry links to the latest version of the model on Microsoft Foundry.
-
+## Deploy a model
+
+```python
+from azure.identity import DefaultAzureCredential
+from azure.mgmt.cognitiveservices import CognitiveServicesManagementClient
+
+SUBSCRIPTION_ID = ""
+RESOURCE_GROUP = ""
+ACCOUNT_NAME = ""
+DEPLOYMENT_NAME = "qwen--qwen3-8-27b"
+
+client = CognitiveServicesManagementClient(
+ credential=DefaultAzureCredential(),
+ subscription_id=SUBSCRIPTION_ID,
+)
-Once a deployment is ready, you can call it through the Foundry endpoint with a standard OpenAI-compatible client:
+deployment = client.managed_compute_deployments.begin_create_or_update(
+ resource_group_name=RESOURCE_GROUP,
+ account_name=ACCOUNT_NAME,
+ deployment_name=DEPLOYMENT_NAME,
+ resource={
+ "sku": {"name": "GlobalManagedCompute", "capacity": 1},
+ "properties": {
+ "model": "azureml://registries/azure-huggingface/models/qwen--qwen3.8-27b/versions/2",
+ "deploymentTemplate": "azureml://registries/azure-huggingface/deploymenttemplates/qwen--qwen3-8-27b--256k-nvidia-h100/labels/latest",
+ "acceleratorType": "H100_80GB",
+ "versionUpgradeOption": "OnceNewDefaultVersionAvailable",
+ },
+ },
+).result()
+
+print(deployment.properties.provisioning_state)
+```
+
+## Run inference
+
+Once the deployment reports `Succeeded`, call it with the OpenAI SDK. The `model` field takes the **deployment name**, not the model name from the catalog:
```python
from openai import OpenAI
client = OpenAI(
- base_url="https://.services.ai.azure.com/openai/v1/",
- api_key="",
+ base_url=f"https://{ACCOUNT_NAME}.services.ai.azure.com/openai/v1",
+ api_key="",
)
response = client.chat.completions.create(
- model="",
+ model=DEPLOYMENT_NAME,
messages=[{"role": "user", "content": "Hello from Microsoft Foundry."}],
)
-```
-## Get started
-
-Before deploying a model from Microsoft Foundry, you will usually need:
-
-- An Azure subscription
-- A Microsoft Foundry project
-- Permission to create deployments in that project
-- Enough quota in the target region for the hardware you want to use
+print(response.choices[0].message.content)
+```
-Follow this path:
+## Next steps
-1. Follow the [quickstart](./get-started/quickstart) to deploy from the Foundry portal or with Python.
-2. Start with [GLM 5.2 FP8](./large-language-models/glm-5-2-fp8) or [DeepSeek V4 Flash 0731](./large-language-models/deepseek-v4-flash-0731).
-3. Verify the deployment through its OpenAI-compatible endpoint.
-4. Connect it to an agent harness, such as [Codex](./harness/codex) or [Hermes Agent](./harness/hermes-agent).
+- Deploy your first model from the portal or from Python in the [quickstart](./get-started/quickstart).
+- Turn reasoning on or off per request in [Enable and disable reasoning](./advanced/reasoning).
+- Connect a deployment to an agent harness such as [Codex](./harness/codex) or [Hermes Agent](./harness/hermes-agent).
From d3bc69bfedc7fd14746edea985bb6b87983d528b Mon Sep 17 00:00:00 2001
From: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
Date: Wed, 19 Aug 2026 13:34:54 +0200
Subject: [PATCH 14/17] Extend catalog, use GitHub-like tips, and more (WIP)
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/_toctree.yml | 4 +
docs/source/advanced/reasoning.mdx | 5 +-
docs/source/catalog/deepseek-ai.mdx | 32 +-
docs/source/catalog/microsoft.mdx | 112 +++
docs/source/catalog/nvidia.mdx | 453 ++++++++++++
docs/source/catalog/qwen.mdx | 653 ++++++++++++------
docs/source/get-started/quickstart.mdx | 14 +-
docs/source/harness/codex.mdx | 15 +-
docs/source/harness/hermes-agent.mdx | 18 +-
docs/source/harness/index.mdx | 47 +-
docs/source/index.mdx | 2 +-
.../deepseek-v4-flash-0731.mdx | 9 +-
.../large-language-models/glm-5-2-fp8.mdx | 7 +-
13 files changed, 1067 insertions(+), 304 deletions(-)
create mode 100644 docs/source/catalog/microsoft.mdx
create mode 100644 docs/source/catalog/nvidia.mdx
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
index 4a82cd0..960b3e9 100644
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -7,6 +7,10 @@
title: Qwen
- local: catalog/deepseek-ai
title: DeepSeek AI
+ - local: catalog/nvidia
+ title: NVIDIA
+ - local: catalog/microsoft
+ title: Microsoft
title: Catalog
- sections:
- local: get-started/quickstart
diff --git a/docs/source/advanced/reasoning.mdx b/docs/source/advanced/reasoning.mdx
index 27fed92..8b63cc7 100644
--- a/docs/source/advanced/reasoning.mdx
+++ b/docs/source/advanced/reasoning.mdx
@@ -55,9 +55,8 @@ response = client.responses.create(
print(response.output_text)
```
-
- Agent harnesses usually set the reasoning effort themselves, so configure it there rather than per request. See [Codex](../harness/codex) or [Hermes Agent](../harness/hermes-agent).
-
+> [!TIP]
+> Agent harnesses usually set the reasoning effort themselves, so configure it there rather than per request. See [Codex](../harness/codex) or [Hermes Agent](../harness/hermes-agent).
## Models without a switch
diff --git a/docs/source/catalog/deepseek-ai.mdx b/docs/source/catalog/deepseek-ai.mdx
index b535106..d313745 100644
--- a/docs/source/catalog/deepseek-ai.mdx
+++ b/docs/source/catalog/deepseek-ai.mdx
@@ -1,30 +1,38 @@
# DeepSeek AI
-[DeepSeek AI](https://huggingface.co/deepseek-ai) builds open mixture-of-experts models for reasoning, coding, and agentic workloads, with long context windows and efficient inference.
-
-Every card links to the latest version of the model in the `azure-huggingface` registry on Microsoft Foundry, and shows its input and output modalities:
+DeepSeek AI is a Chinese AI research company that releases open mixture-of-experts models built for reasoning, coding, and agentic work over long contexts.
+ TText
+
+
+
diff --git a/docs/source/catalog/qwen.mdx b/docs/source/catalog/qwen.mdx
new file mode 100644
index 0000000..50c6630
--- /dev/null
+++ b/docs/source/catalog/qwen.mdx
@@ -0,0 +1,531 @@
+# Qwen
+
+[Qwen](https://huggingface.co/Qwen) is Alibaba Cloud's open model family, spanning dense and mixture-of-experts text models, vision-language models, embeddings, and image generation. It is the largest family in the Hugging Face collection on Microsoft Foundry.
+
+Every card links to the latest version of the model in the `azure-huggingface` registry on Microsoft Foundry, and shows its input and output modalities:
+
+
+ TText
+ Image
+ Video
+ Embeddings
+
+
+## Chat completion
+
+
+
+ 
+
+ Qwen3-0.6B
+ qwen--qwen3-0.6b · v1
+ Compact Qwen3 0.6B chat model for lightweight text generation.
+ T→T
+
+
+
+
+
+ Qwen3-14B
+ qwen--qwen3-14b · v2
+ Dense 14.8B Qwen3 reasoning model with multilingual chat, tool use, and switchable thinking modes.
+ T→T
+
+
+
+
+
+ Qwen3-30B-A3B-Instruct-2507
+ qwen--qwen3-30b-a3b-instruct-2507 · v1
+ Qwen3 30B-A3B instruction model for long-context reasoning and generation.
+ T→T
+
+
+
+
+
+ Qwen3-30B-A3B-Instruct-2507-FP8
+ qwen--qwen3-30b-a3b-instruct-2507-fp8 · v1
+ FP8 Qwen3 30B-A3B instruct model for long-context reasoning and generation.
+ T→T
+
+
+
+
+
+ Qwen3-30B-A3B-Thinking-2507
+ qwen--qwen3-30b-a3b-thinking-2507 · v1
+ Qwen3 30B-A3B thinking model for long-context reasoning and generation.
+ T→T
+
+
+
+
+
+ Qwen3-32B
+ qwen--qwen3-32b · v3
+ Dense 32.8B Qwen3 reasoning model with switchable thinking modes, multilingual chat, and agentic workflows.
+ T→T
+
+
+
+
+
+ Qwen3-4B-Instruct-2507
+ qwen--qwen3-4b-instruct-2507 · v1
+ Qwen3 4B instruction model for compact long-context chat and text generation.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct
+ qwen--qwen3-coder-30b-a3b-instruct · v1
+ Qwen3 Coder 30B-A3B instruction model for code generation and agentic workflows.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct-FP8
+ qwen--qwen3-coder-30b-a3b-instruct-fp8 · v1
+ FP8 Qwen3 Coder 30B-A3B model for code generation and agentic workflows.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-Next
+ qwen--qwen3-coder-next · v1
+ Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-Next-FP8
+ qwen--qwen3-coder-next-fp8 · v1
+ FP8 Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
+ T→T
+
+
+
+
+
+ Qwen3-VL-30B-A3B-Instruct
+ qwen--qwen3-vl-30b-a3b-instruct · v1
+ Qwen3-VL 30B-A3B multimodal model for image, video, and text generation.
+ T→T
+
+
+
+
+
+ Qwen3-VL-30B-A3B-Instruct-FP8
+ qwen--qwen3-vl-30b-a3b-instruct-fp8 · v1
+ FP8 Qwen3-VL 30B-A3B instruct model for multimodal image, video, and text generation.
+ T→T
+
+
+
+
+
+ Qwen3.5-0.8B
+ qwen--qwen3.5-0.8b · v1
+ Compact multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.5-2B
+ qwen--qwen3.5-2b · v1
+ Small multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.5-35B-A3B
+ qwen--qwen3.5-35b-a3b · v1
+ Qwen3.5 35B-A3B multimodal model for visual reasoning and long-context generation.
+ T→T
+
+
+
+
+
+ Qwen3.5-397B-A17B
+ qwen--qwen3.5-397b-a17b · v1
+ Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and long-context video.
+ T→T
+
+
+
+
+
+ Qwen3.5-397B-A17B-FP8
+ qwen--qwen3.5-397b-a17b-fp8 · v1
+ FP8 Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and video.
+ T→T
+
+
+
+
+
+ Qwen3.5-4B
+ qwen--qwen3.5-4b · v1
+ Qwen3.5 4B multimodal model for visual reasoning, OCR, and long-context generation.
+ T→T
+
+
+
+
+
+ Qwen3.5-9B
+ qwen--qwen3.5-9b · v1
+ Qwen3.5 9B multimodal model for visual reasoning, OCR, and long-context generation.
+ T→T
+
+
+
+
+
+ Qwen3.6-27B
+ qwen--qwen3.6-27b · v2
+ Multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.6-27B-FP8
+ qwen--qwen3.6-27b-fp8 · v5
+ FP8-quantized multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.8-27B
+ qwen--qwen3.8-27b · v2
+ Compact Qwen3.8 27B vision-language model for coding, agents, reasoning, images, and video.
+ T→T
+
+
+
+
+
+ QwQ-32B
+ qwen--qwq-32b · v1
+ QwQ 32B reasoning model for math, coding, and analytical text generation.
+ T→T
+
+
+
+
+
+ QwQ-32B-Preview
+ qwen--qwq-32b-preview · v2
+ Preview QwQ 32B reasoning model for experimental math and coding tasks.
+ T→T
+
+
+
+
+## Embeddings
+
+
+
+ Qwen3-0.6B
+ qwen--qwen3-0.6b · v1
+ Compact Qwen3 0.6B chat model for lightweight text generation.
+ T→T
+
+
+
+
+
+ Qwen3-14B
+ qwen--qwen3-14b · v2
+ Dense 14.8B Qwen3 reasoning model with multilingual chat, tool use, and switchable thinking modes.
+ T→T
+
+
+
+
+
+ Qwen3-30B-A3B-Instruct-2507
+ qwen--qwen3-30b-a3b-instruct-2507 · v1
+ Qwen3 30B-A3B instruction model for long-context reasoning and generation.
+ T→T
+
+
+
+
+
+ Qwen3-30B-A3B-Instruct-2507-FP8
+ qwen--qwen3-30b-a3b-instruct-2507-fp8 · v1
+ FP8 Qwen3 30B-A3B instruct model for long-context reasoning and generation.
+ T→T
+
+
+
+
+
+ Qwen3-30B-A3B-Thinking-2507
+ qwen--qwen3-30b-a3b-thinking-2507 · v1
+ Qwen3 30B-A3B thinking model for long-context reasoning and generation.
+ T→T
+
+
+
+
+
+ Qwen3-32B
+ qwen--qwen3-32b · v3
+ Dense 32.8B Qwen3 reasoning model with switchable thinking modes, multilingual chat, and agentic workflows.
+ T→T
+
+
+
+
+
+ Qwen3-4B-Instruct-2507
+ qwen--qwen3-4b-instruct-2507 · v1
+ Qwen3 4B instruction model for compact long-context chat and text generation.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct
+ qwen--qwen3-coder-30b-a3b-instruct · v1
+ Qwen3 Coder 30B-A3B instruction model for code generation and agentic workflows.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct-FP8
+ qwen--qwen3-coder-30b-a3b-instruct-fp8 · v1
+ FP8 Qwen3 Coder 30B-A3B model for code generation and agentic workflows.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-Next
+ qwen--qwen3-coder-next · v1
+ Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
+ T→T
+
+
+
+
+
+ Qwen3-Coder-Next-FP8
+ qwen--qwen3-coder-next-fp8 · v1
+ FP8 Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
+ T→T
+
+
+
+
+
+ Qwen3-VL-30B-A3B-Instruct
+ qwen--qwen3-vl-30b-a3b-instruct · v1
+ Qwen3-VL 30B-A3B multimodal model for image, video, and text generation.
+ T→T
+
+
+
+
+
+ Qwen3-VL-30B-A3B-Instruct-FP8
+ qwen--qwen3-vl-30b-a3b-instruct-fp8 · v1
+ FP8 Qwen3-VL 30B-A3B instruct model for multimodal image, video, and text generation.
+ T→T
+
+
+
+
+
+ Qwen3.5-0.8B
+ qwen--qwen3.5-0.8b · v1
+ Compact multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.5-2B
+ qwen--qwen3.5-2b · v1
+ Small multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.5-35B-A3B
+ qwen--qwen3.5-35b-a3b · v1
+ Qwen3.5 35B-A3B multimodal model for visual reasoning and long-context generation.
+ T→T
+
+
+
+
+
+ Qwen3.5-397B-A17B
+ qwen--qwen3.5-397b-a17b · v1
+ Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and long-context video.
+ T→T
+
+
+
+
+
+ Qwen3.5-397B-A17B-FP8
+ qwen--qwen3.5-397b-a17b-fp8 · v1
+ FP8 Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and video.
+ T→T
+
+
+
+
+
+ Qwen3.5-4B
+ qwen--qwen3.5-4b · v1
+ Qwen3.5 4B multimodal model for visual reasoning, OCR, and long-context generation.
+ T→T
+
+
+
+
+
+ Qwen3.5-9B
+ qwen--qwen3.5-9b · v1
+ Qwen3.5 9B multimodal model for visual reasoning, OCR, and long-context generation.
+ T→T
+
+
+
+
+
+ Qwen3.6-27B
+ qwen--qwen3.6-27b · v2
+ Multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.6-27B-FP8
+ qwen--qwen3.6-27b-fp8 · v5
+ FP8-quantized multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+ T→T
+
+
+
+
+
+ Qwen3.8-27B
+ qwen--qwen3.8-27b · v2
+ Compact Qwen3.8 27B vision-language model for coding, agents, reasoning, images, and video.
+ T→T
+
+
+
+
+
+ QwQ-32B
+ qwen--qwq-32b · v1
+ QwQ 32B reasoning model for math, coding, and analytical text generation.
+ T→T
+
+
+
+
+
+ QwQ-32B-Preview
+ qwen--qwq-32b-preview · v2
+ Preview QwQ 32B reasoning model for experimental math and coding tasks.
+ T→T
+
+
+
+
+
+
+ Qwen3-Embedding-0.6B
+ qwen--qwen3-embedding-0.6b · v1
+ Compact multilingual embedding model with flexible output dimensions for retrieval, classification, clustering, and code retrieval.
+ T→
+
+
+
+
+
+ Qwen3-Embedding-8B
+ qwen--qwen3-embedding-8b · v2
+ State-of-the-art 8B multilingual embedding model for retrieval, classification, clustering, bitext mining, and code retrieval.
+ T→
+
+
+
+
+## Text to image
+
+
diff --git a/docs/source/get-started/quickstart.mdx b/docs/source/get-started/quickstart.mdx
index 3897a13..1aefcc2 100644
--- a/docs/source/get-started/quickstart.mdx
+++ b/docs/source/get-started/quickstart.mdx
@@ -1,6 +1,6 @@
# Quickstart
-Deploy a Hugging Face model on Microsoft Foundry, test it in the Playground, and call it from an application. You can use the Foundry portal for a guided setup or the Python SDK for a repeatable deployment.
+Deploy a model from the Hugging Face collection on Microsoft Foundry, then call it from Python.
## Before you begin
@@ -8,67 +8,62 @@ You need:
- An Azure subscription and a Microsoft Foundry project
- `Cognitive Services Contributor` access to create the deployment
-- `Foundry User` access to call it from the Playground
-- Managed compute quota for the accelerator used by your model
+- Managed compute quota for the accelerator that your model requires
-Managed compute is currently in preview and uses dedicated GPU capacity. Delete deployments that you no longer need to stop billing.
+
+
+ Qwen3-Embedding-0.6B
+ qwen--qwen3-embedding-0.6b · v1
+ Compact multilingual embedding model with flexible output dimensions for retrieval, classification, clustering, and code retrieval.
+ T→
+
+
+
+
+
+ Qwen3-Embedding-8B
+ qwen--qwen3-embedding-8b · v2
+ State-of-the-art 8B multilingual embedding model for retrieval, classification, clustering, bitext mining, and code retrieval.
+ T→
+
+
+
- TText
+ TText
-
+## Chat completion
+
+
diff --git a/docs/source/catalog/microsoft.mdx b/docs/source/catalog/microsoft.mdx
new file mode 100644
index 0000000..23fea6c
--- /dev/null
+++ b/docs/source/catalog/microsoft.mdx
@@ -0,0 +1,112 @@
+# Microsoft
+
+Microsoft releases the Phi family of small language models, along with domain-tuned variants such as MediPhi for clinical text.
+
+
+ TText
+
+
+## Chat completion
+
+
+
+ 
+
+
+
+ OptiMind-SFT
+
+
+
+ Specialized 20B parameter model designed to bridge the gap between natural language and executable optimization solvers.
+
+ T→T
+ Dec 2025
+
+
+
+
+
+
+
+
+ MediPhi-Instruct
+
+
+
+ Instruction-aligned MediPhi model for clinical and biomedical NLP tasks.
+
+ T→T
+ Jul 2025
+
+
+
+
+
+
+
+
+ MediPhi
+
+
+
+ Phi-3.5 based clinical language model for biomedical and medical NLP workflows.
+
+ T→T
+ Feb 2025
+
+
+
+
+
+
+
+
+ phi-4
+
+
+
+ 14B dense Phi model for English chat, reasoning, math, and code generation.
+
+ T→T
+ Dec 2024
+
+
+
+
diff --git a/docs/source/catalog/nvidia.mdx b/docs/source/catalog/nvidia.mdx
new file mode 100644
index 0000000..e99339b
--- /dev/null
+++ b/docs/source/catalog/nvidia.mdx
@@ -0,0 +1,453 @@
+# NVIDIA
+
+NVIDIA releases the Nemotron and Cosmos families, together with FP8 and NVFP4 quantizations of them that lower the accelerator footprint needed to serve each model.
+
+
+
+
+
+ OptiMind-SFT
+
+
+
+ Specialized 20B parameter model designed to bridge the gap between natural language and executable optimization solvers.
+
+ T→T
+ Dec 2025
+
+
+
+
+
+
+
+
+ MediPhi-Instruct
+
+
+
+ Instruction-aligned MediPhi model for clinical and biomedical NLP tasks.
+
+ T→T
+ Jul 2025
+
+
+
+
+
+
+
+
+ MediPhi
+
+
+
+ Phi-3.5 based clinical language model for biomedical and medical NLP workflows.
+
+ T→T
+ Feb 2025
+
+
+
+
+
+
+
+
+ phi-4
+
+
+
+ 14B dense Phi model for English chat, reasoning, math, and code generation.
+
+ T→T
+ Dec 2024
+
+
+
+
+ TText
+ Image
+ Video
+ Audio
+ Embeddings
+
+
+## Chat completion
+
+
+
+ 
+
+
+
+ NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
+ New
+
+
+
+ Efficient Nemotron 3.5 Lightning 30B NVFP4 model for long-running agents and local inference.
+
+ T→T
+ Aug 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
+ New
+
+
+
+ Full-precision Nemotron 3.5 Lightning 30B reference model for customization and domain adaptation.
+
+ T→T
+ Aug 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4
+
+
+
+ Frontier-scale NVFP4 reasoning and chat model for complex agentic workflows, long-context analysis, tool use, multilingual reasoning, and high-stakes RAG.
+
+ T→T
+ Jun 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3-Super-120B-A12B-FP8
+
+
+
+ Instruction-tuned reasoning model for chat, tool use, and complex problem solving.
+
+ T→T
+ Mar 2026
+
+
+
+
+
+
+
+
+ Nemotron-Terminal-32B
+
+
+
+ NVIDIA Qwen3-based 32B model specialized for autonomous terminal interaction.
+
+ T→T
+ Feb 2026
+
+
+
+
+
+
+
+
+ Nemotron-Terminal-14B
+
+
+
+ NVIDIA Qwen3-based 14B model specialized for autonomous terminal interaction.
+
+ T→T
+ Feb 2026
+
+
+
+
+
+
+
+
+ Nemotron-Terminal-8B
+
+
+
+ NVIDIA Qwen3-based 8B model specialized for autonomous terminal interaction.
+
+ T→T
+ Feb 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3-Nano-30B-A3B-FP8
+
+
+
+ Instruction-tuned reasoning model for chat, tool use, and complex problem solving.
+
+ T→T
+ Dec 2025
+
+
+
+
+
+
+
+
+ OpenCodeReasoning-Nemotron-1.1-32B
+
+
+
+ NVIDIA 32B code-reasoning model derived from Qwen2.5-32B-Instruct.
+
+ T→T
+ Jun 2025
+
+
+
+
+
+
+
+
+ OpenCodeReasoning-Nemotron-1.1-14B
+
+
+
+ NVIDIA 14B code-reasoning model derived from Qwen2.5-14B-Instruct.
+
+ T→T
+ Jun 2025
+
+
+
+
+
+
+
+
+ Cosmos-Reason1-7B
+
+
+
+ VLM for physical AI reasoning over text, images, and video.
+
+ T→T
+ Apr 2025
+
+
+
+
+
+## Multimodal generation
+
+
+
+
+
+ NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
+ New
+
+
+
+ Efficient Nemotron 3.5 Lightning 30B NVFP4 model for long-running agents and local inference.
+
+ T→T
+ Aug 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
+ New
+
+
+
+ Full-precision Nemotron 3.5 Lightning 30B reference model for customization and domain adaptation.
+
+ T→T
+ Aug 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4
+
+
+
+ Frontier-scale NVFP4 reasoning and chat model for complex agentic workflows, long-context analysis, tool use, multilingual reasoning, and high-stakes RAG.
+
+ T→T
+ Jun 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3-Super-120B-A12B-FP8
+
+
+
+ Instruction-tuned reasoning model for chat, tool use, and complex problem solving.
+
+ T→T
+ Mar 2026
+
+
+
+
+
+
+
+
+ Nemotron-Terminal-32B
+
+
+
+ NVIDIA Qwen3-based 32B model specialized for autonomous terminal interaction.
+
+ T→T
+ Feb 2026
+
+
+
+
+
+
+
+
+ Nemotron-Terminal-14B
+
+
+
+ NVIDIA Qwen3-based 14B model specialized for autonomous terminal interaction.
+
+ T→T
+ Feb 2026
+
+
+
+
+
+
+
+
+ Nemotron-Terminal-8B
+
+
+
+ NVIDIA Qwen3-based 8B model specialized for autonomous terminal interaction.
+
+ T→T
+ Feb 2026
+
+
+
+
+
+
+
+
+ NVIDIA-Nemotron-3-Nano-30B-A3B-FP8
+
+
+
+ Instruction-tuned reasoning model for chat, tool use, and complex problem solving.
+
+ T→T
+ Dec 2025
+
+
+
+
+
+
+
+
+ OpenCodeReasoning-Nemotron-1.1-32B
+
+
+
+ NVIDIA 32B code-reasoning model derived from Qwen2.5-32B-Instruct.
+
+ T→T
+ Jun 2025
+
+
+
+
+
+
+
+
+ OpenCodeReasoning-Nemotron-1.1-14B
+
+
+
+ NVIDIA 14B code-reasoning model derived from Qwen2.5-14B-Instruct.
+
+ T→T
+ Jun 2025
+
+
+
+
+
+
+
+
+ Cosmos-Reason1-7B
+
+
+
+ VLM for physical AI reasoning over text, images, and video.
+
+ T→T
+ Apr 2025
+
+
+
+
+
+
+
+
+
+ Cosmos3-Edge
+
+
+
+ Cosmos3 Edge omni model for text, image, video, audio, and action generation.
+
+ T→T
+ Jul 2026
+
+
+
+
+
+
+
+
+ Cosmos3-Nano
+
+
+
+ Cosmos3 Nano omni model for text, image, video, audio, and action generation.
+
+ T→T
+ Mar 2026
+
+
+
+
+
+
+
+
+ Cosmos3-Super
+
+
+
+ Cosmos3 Super omni model for text, image, video, audio, and action generation.
+
+ T→T
+ Mar 2026
+
+
+
+
+
+## Embeddings
+
+
+
+
+
+ Cosmos3-Edge
+
+
+
+ Cosmos3 Edge omni model for text, image, video, audio, and action generation.
+
+ T→T
+ Jul 2026
+
+
+
+
+
+
+
+
+ Cosmos3-Nano
+
+
+
+ Cosmos3 Nano omni model for text, image, video, audio, and action generation.
+
+ T→T
+ Mar 2026
+
+
+
+
+
+
+
+
+ Cosmos3-Super
+
+
+
+ Cosmos3 Super omni model for text, image, video, audio, and action generation.
+
+ T→T
+ Mar 2026
+
+
+
+
+
+
+
+
+
+ Nemotron-3-Embed-1B-NVFP4
+
+
+
+ NVFP4-quantized multilingual 1B embedding model for retrieval, semantic search, and RAG.
+
+ T→
+ Jul 2026
+
+
+
+
+
+
+
+
+ Nemotron-3-Embed-8B-BF16
+
+
+
+ Multilingual 8B embedding model for retrieval, semantic search, and RAG with flexible 4096-dim vectors.
+
+ T→
+ Jul 2026
+
+
+
+
+
+
+
+
+ Nemotron-3-Embed-1B-BF16
+
+
+
+ Multilingual 1B embedding model for retrieval, semantic search, and RAG with flexible 2048-dim vectors.
+
+ T→
+ Jul 2026
+
+
+
+
diff --git a/docs/source/catalog/qwen.mdx b/docs/source/catalog/qwen.mdx
index 50c6630..90f8686 100644
--- a/docs/source/catalog/qwen.mdx
+++ b/docs/source/catalog/qwen.mdx
@@ -1,531 +1,726 @@
# Qwen
-[Qwen](https://huggingface.co/Qwen) is Alibaba Cloud's open model family, spanning dense and mixture-of-experts text models, vision-language models, embeddings, and image generation. It is the largest family in the Hugging Face collection on Microsoft Foundry.
-
-Every card links to the latest version of the model in the `azure-huggingface` registry on Microsoft Foundry, and shows its input and output modalities:
+Qwen is the open model family from Alibaba Cloud, released under permissive licenses and covering text, vision-language, coding, embedding, and image generation models.
+
+
+
+ Nemotron-3-Embed-1B-NVFP4
+
+
+
+ NVFP4-quantized multilingual 1B embedding model for retrieval, semantic search, and RAG.
+
+ T→
+ Jul 2026
+
+
+
+
+
+
+
+
+ Nemotron-3-Embed-8B-BF16
+
+
+
+ Multilingual 8B embedding model for retrieval, semantic search, and RAG with flexible 4096-dim vectors.
+
+ T→
+ Jul 2026
+
+
+
+
+
+
+
+
+ Nemotron-3-Embed-1B-BF16
+
+
+
+ Multilingual 1B embedding model for retrieval, semantic search, and RAG with flexible 2048-dim vectors.
+
+ T→
+ Jul 2026
+
+
+
+
- TText
- Image
- Video
- Embeddings
+ TText
+ Image
+ Video
+ Embeddings
## Chat completion
-
+
-
- Qwen3-0.6B
- qwen--qwen3-0.6b · v1
- Compact Qwen3 0.6B chat model for lightweight text generation.
- T→T
+
+
+
+ Qwen3.8-27B
+ New
+
+
+
+ Compact Qwen3.8 27B vision-language model for coding, agents, reasoning, images, and video.
+
+ T→T
+ Aug 2026
+
-
- Qwen3-14B
- qwen--qwen3-14b · v2
- Dense 14.8B Qwen3 reasoning model with multilingual chat, tool use, and switchable thinking modes.
- T→T
+
+
+
+ Qwen3.6-27B-FP8
+
+
+
+ FP8-quantized multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Apr 2026
+
-
- Qwen3-30B-A3B-Instruct-2507
- qwen--qwen3-30b-a3b-instruct-2507 · v1
- Qwen3 30B-A3B instruction model for long-context reasoning and generation.
- T→T
+
+
+
+ Qwen3.6-27B
+
+
+
+ Multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Apr 2026
+
-
- Qwen3-30B-A3B-Instruct-2507-FP8
- qwen--qwen3-30b-a3b-instruct-2507-fp8 · v1
- FP8 Qwen3 30B-A3B instruct model for long-context reasoning and generation.
- T→T
+
+
+
+ Qwen3.5-0.8B
+
+
+
+ Compact multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-30B-A3B-Thinking-2507
- qwen--qwen3-30b-a3b-thinking-2507 · v1
- Qwen3 30B-A3B thinking model for long-context reasoning and generation.
- T→T
+
+
+
+ Qwen3.5-2B
+
+
+
+ Small multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-32B
- qwen--qwen3-32b · v3
- Dense 32.8B Qwen3 reasoning model with switchable thinking modes, multilingual chat, and agentic workflows.
- T→T
+
+
+
+ Qwen3.5-4B
+
+
+
+ Qwen3.5 4B multimodal model for visual reasoning, OCR, and long-context generation.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-4B-Instruct-2507
- qwen--qwen3-4b-instruct-2507 · v1
- Qwen3 4B instruction model for compact long-context chat and text generation.
- T→T
+
+
+
+ Qwen3.5-9B
+
+
+
+ Qwen3.5 9B multimodal model for visual reasoning, OCR, and long-context generation.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-30B-A3B-Instruct
- qwen--qwen3-coder-30b-a3b-instruct · v1
- Qwen3 Coder 30B-A3B instruction model for code generation and agentic workflows.
- T→T
+
+
+
+ Qwen3.5-35B-A3B
+
+
+
+ Qwen3.5 35B-A3B multimodal model for visual reasoning and long-context generation.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-30B-A3B-Instruct-FP8
- qwen--qwen3-coder-30b-a3b-instruct-fp8 · v1
- FP8 Qwen3 Coder 30B-A3B model for code generation and agentic workflows.
- T→T
+
+
+
+ Qwen3.5-397B-A17B-FP8
+
+
+
+ FP8 Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and video.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-Next
- qwen--qwen3-coder-next · v1
- Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
- T→T
+
+
+
+ Qwen3.5-397B-A17B
+
+
+
+ Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and long-context video.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-Next-FP8
- qwen--qwen3-coder-next-fp8 · v1
+
+
+
+ Qwen3-Coder-Next-FP8
+
+
+
FP8 Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
- T→T
+
+ T→T
+ Feb 2026
+
-
- Qwen3-VL-30B-A3B-Instruct
- qwen--qwen3-vl-30b-a3b-instruct · v1
- Qwen3-VL 30B-A3B multimodal model for image, video, and text generation.
- T→T
+
+
+
+ Qwen3-Coder-Next
+
+
+
+ Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
+
+ T→T
+ Jan 2026
+
-
- Qwen3-VL-30B-A3B-Instruct-FP8
- qwen--qwen3-vl-30b-a3b-instruct-fp8 · v1
+
+
+
+ Qwen3-VL-30B-A3B-Instruct-FP8
+
+
+
FP8 Qwen3-VL 30B-A3B instruct model for multimodal image, video, and text generation.
- T→T
+
+ T→T
+ Oct 2025
+
-
- Qwen3.5-0.8B
- qwen--qwen3.5-0.8b · v1
- Compact multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-VL-30B-A3B-Instruct
+
+
+
+ Qwen3-VL 30B-A3B multimodal model for image, video, and text generation.
+
+ T→T
+ Sep 2025
+
-
- Qwen3.5-2B
- qwen--qwen3.5-2b · v1
- Small multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-4B-Instruct-2507
+
+
+
+ Qwen3 4B instruction model for compact long-context chat and text generation.
+
+ T→T
+ Aug 2025
+
-
- Qwen3.5-35B-A3B
- qwen--qwen3.5-35b-a3b · v1
- Qwen3.5 35B-A3B multimodal model for visual reasoning and long-context generation.
- T→T
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct-FP8
+
+
+
+ FP8 Qwen3 Coder 30B-A3B model for code generation and agentic workflows.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-397B-A17B
- qwen--qwen3.5-397b-a17b · v1
- Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and long-context video.
- T→T
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct
+
+
+
+ Qwen3 Coder 30B-A3B instruction model for code generation and agentic workflows.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-397B-A17B-FP8
- qwen--qwen3.5-397b-a17b-fp8 · v1
- FP8 Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and video.
- T→T
+
+
+
+ Qwen3-30B-A3B-Thinking-2507
+
+
+
+ Qwen3 30B-A3B thinking model for long-context reasoning and generation.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-4B
- qwen--qwen3.5-4b · v1
- Qwen3.5 4B multimodal model for visual reasoning, OCR, and long-context generation.
- T→T
+
+
+
+ Qwen3-30B-A3B-Instruct-2507-FP8
+
+
+
+ FP8 Qwen3 30B-A3B instruct model for long-context reasoning and generation.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-9B
- qwen--qwen3.5-9b · v1
- Qwen3.5 9B multimodal model for visual reasoning, OCR, and long-context generation.
- T→T
+
+
+
+ Qwen3-30B-A3B-Instruct-2507
+
+
+
+ Qwen3 30B-A3B instruction model for long-context reasoning and generation.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.6-27B
- qwen--qwen3.6-27b · v2
- Multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-32B
+
+
+
+ Dense 32.8B Qwen3 reasoning model with switchable thinking modes, multilingual chat, and agentic workflows.
+
+ T→T
+ Apr 2025
+
-
- Qwen3.6-27B-FP8
- qwen--qwen3.6-27b-fp8 · v5
- FP8-quantized multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-14B
+
+
+
+ Dense 14.8B Qwen3 reasoning model with multilingual chat, tool use, and switchable thinking modes.
+
+ T→T
+ Apr 2025
+
-
- Qwen3.8-27B
- qwen--qwen3.8-27b · v2
- Compact Qwen3.8 27B vision-language model for coding, agents, reasoning, images, and video.
- T→T
+
+
+
+ Qwen3-0.6B
+
+
+
+ Compact Qwen3 0.6B chat model for lightweight text generation.
+
+ T→T
+ Apr 2025
+
-
- QwQ-32B
- qwen--qwq-32b · v1
+
+
+
+ QwQ-32B
+
+
+
QwQ 32B reasoning model for math, coding, and analytical text generation.
- T→T
+
+ T→T
+ Mar 2025
+
-
- QwQ-32B-Preview
- qwen--qwq-32b-preview · v2
+
+
+
+ QwQ-32B-Preview
+
+
+
Preview QwQ 32B reasoning model for experimental math and coding tasks.
- T→T
+
+ T→T
+ Nov 2024
+
-## Embeddings
+## Text to image
-
-
- Qwen3-0.6B
- qwen--qwen3-0.6b · v1
- Compact Qwen3 0.6B chat model for lightweight text generation.
- T→T
+
+
+
+ Qwen3.8-27B
+ New
+
+
+
+ Compact Qwen3.8 27B vision-language model for coding, agents, reasoning, images, and video.
+
+ T→T
+ Aug 2026
+
-
- Qwen3-14B
- qwen--qwen3-14b · v2
- Dense 14.8B Qwen3 reasoning model with multilingual chat, tool use, and switchable thinking modes.
- T→T
+
+
+
+ Qwen3.6-27B-FP8
+
+
+
+ FP8-quantized multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Apr 2026
+
-
- Qwen3-30B-A3B-Instruct-2507
- qwen--qwen3-30b-a3b-instruct-2507 · v1
- Qwen3 30B-A3B instruction model for long-context reasoning and generation.
- T→T
+
+
+
+ Qwen3.6-27B
+
+
+
+ Multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Apr 2026
+
-
- Qwen3-30B-A3B-Instruct-2507-FP8
- qwen--qwen3-30b-a3b-instruct-2507-fp8 · v1
- FP8 Qwen3 30B-A3B instruct model for long-context reasoning and generation.
- T→T
+
+
+
+ Qwen3.5-0.8B
+
+
+
+ Compact multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-30B-A3B-Thinking-2507
- qwen--qwen3-30b-a3b-thinking-2507 · v1
- Qwen3 30B-A3B thinking model for long-context reasoning and generation.
- T→T
+
+
+
+ Qwen3.5-2B
+
+
+
+ Small multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-32B
- qwen--qwen3-32b · v3
- Dense 32.8B Qwen3 reasoning model with switchable thinking modes, multilingual chat, and agentic workflows.
- T→T
+
+
+
+ Qwen3.5-4B
+
+
+
+ Qwen3.5 4B multimodal model for visual reasoning, OCR, and long-context generation.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-4B-Instruct-2507
- qwen--qwen3-4b-instruct-2507 · v1
- Qwen3 4B instruction model for compact long-context chat and text generation.
- T→T
+
+
+
+ Qwen3.5-9B
+
+
+
+ Qwen3.5 9B multimodal model for visual reasoning, OCR, and long-context generation.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-30B-A3B-Instruct
- qwen--qwen3-coder-30b-a3b-instruct · v1
- Qwen3 Coder 30B-A3B instruction model for code generation and agentic workflows.
- T→T
+
+
+
+ Qwen3.5-35B-A3B
+
+
+
+ Qwen3.5 35B-A3B multimodal model for visual reasoning and long-context generation.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-30B-A3B-Instruct-FP8
- qwen--qwen3-coder-30b-a3b-instruct-fp8 · v1
- FP8 Qwen3 Coder 30B-A3B model for code generation and agentic workflows.
- T→T
+
+
+
+ Qwen3.5-397B-A17B-FP8
+
+
+
+ FP8 Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and video.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-Next
- qwen--qwen3-coder-next · v1
- Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
- T→T
+
+
+
+ Qwen3.5-397B-A17B
+
+
+
+ Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and long-context video.
+
+ T→T
+ Feb 2026
+
-
- Qwen3-Coder-Next-FP8
- qwen--qwen3-coder-next-fp8 · v1
+
+
+
+ Qwen3-Coder-Next-FP8
+
+
+
FP8 Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
- T→T
+
+ T→T
+ Feb 2026
+
-
- Qwen3-VL-30B-A3B-Instruct
- qwen--qwen3-vl-30b-a3b-instruct · v1
- Qwen3-VL 30B-A3B multimodal model for image, video, and text generation.
- T→T
+
+
+
+ Qwen3-Coder-Next
+
+
+
+ Qwen3-Coder-Next model for agentic coding, tool use, and long-context software workflows.
+
+ T→T
+ Jan 2026
+
-
- Qwen3-VL-30B-A3B-Instruct-FP8
- qwen--qwen3-vl-30b-a3b-instruct-fp8 · v1
+
+
+
+ Qwen3-VL-30B-A3B-Instruct-FP8
+
+
+
FP8 Qwen3-VL 30B-A3B instruct model for multimodal image, video, and text generation.
- T→T
+
+ T→T
+ Oct 2025
+
-
- Qwen3.5-0.8B
- qwen--qwen3.5-0.8b · v1
- Compact multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-VL-30B-A3B-Instruct
+
+
+
+ Qwen3-VL 30B-A3B multimodal model for image, video, and text generation.
+
+ T→T
+ Sep 2025
+
-
- Qwen3.5-2B
- qwen--qwen3.5-2b · v1
- Small multimodal Qwen3.5 model for visual reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-4B-Instruct-2507
+
+
+
+ Qwen3 4B instruction model for compact long-context chat and text generation.
+
+ T→T
+ Aug 2025
+
-
- Qwen3.5-35B-A3B
- qwen--qwen3.5-35b-a3b · v1
- Qwen3.5 35B-A3B multimodal model for visual reasoning and long-context generation.
- T→T
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct-FP8
+
+
+
+ FP8 Qwen3 Coder 30B-A3B model for code generation and agentic workflows.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-397B-A17B
- qwen--qwen3.5-397b-a17b · v1
- Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and long-context video.
- T→T
+
+
+
+ Qwen3-Coder-30B-A3B-Instruct
+
+
+
+ Qwen3 Coder 30B-A3B instruction model for code generation and agentic workflows.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-397B-A17B-FP8
- qwen--qwen3.5-397b-a17b-fp8 · v1
- FP8 Qwen3.5 397B-A17B multimodal MoE for reasoning, coding, agents, OCR, GUI, and video.
- T→T
+
+
+
+ Qwen3-30B-A3B-Thinking-2507
+
+
+
+ Qwen3 30B-A3B thinking model for long-context reasoning and generation.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-4B
- qwen--qwen3.5-4b · v1
- Qwen3.5 4B multimodal model for visual reasoning, OCR, and long-context generation.
- T→T
+
+
+
+ Qwen3-30B-A3B-Instruct-2507-FP8
+
+
+
+ FP8 Qwen3 30B-A3B instruct model for long-context reasoning and generation.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.5-9B
- qwen--qwen3.5-9b · v1
- Qwen3.5 9B multimodal model for visual reasoning, OCR, and long-context generation.
- T→T
+
+
+
+ Qwen3-30B-A3B-Instruct-2507
+
+
+
+ Qwen3 30B-A3B instruction model for long-context reasoning and generation.
+
+ T→T
+ Jul 2025
+
-
- Qwen3.6-27B
- qwen--qwen3.6-27b · v2
- Multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-32B
+
+
+
+ Dense 32.8B Qwen3 reasoning model with switchable thinking modes, multilingual chat, and agentic workflows.
+
+ T→T
+ Apr 2025
+
-
- Qwen3.6-27B-FP8
- qwen--qwen3.6-27b-fp8 · v5
- FP8-quantized multimodal 27B Qwen3.6 model for coding, reasoning, OCR, GUI interaction, and long-context video understanding.
- T→T
+
+
+
+ Qwen3-14B
+
+
+
+ Dense 14.8B Qwen3 reasoning model with multilingual chat, tool use, and switchable thinking modes.
+
+ T→T
+ Apr 2025
+
-
- Qwen3.8-27B
- qwen--qwen3.8-27b · v2
- Compact Qwen3.8 27B vision-language model for coding, agents, reasoning, images, and video.
- T→T
+
+
+
+ Qwen3-0.6B
+
+
+
+ Compact Qwen3 0.6B chat model for lightweight text generation.
+
+ T→T
+ Apr 2025
+
-
- QwQ-32B
- qwen--qwq-32b · v1
+
+
+
+ QwQ-32B
+
+
+
QwQ 32B reasoning model for math, coding, and analytical text generation.
- T→T
+
+ T→T
+ Mar 2025
+
-
- QwQ-32B-Preview
- qwen--qwq-32b-preview · v2
+
+
+
+ QwQ-32B-Preview
+
+
+
Preview QwQ 32B reasoning model for experimental math and coding tasks.
- T→T
+
+ T→T
+ Nov 2024
+
+
+
+## Embeddings
+
+
-
-## Text to image
-
-
diff --git a/docs/source/get-started/quickstart.mdx b/docs/source/get-started/quickstart.mdx
index 1aefcc2..52b0f7f 100644
--- a/docs/source/get-started/quickstart.mdx
+++ b/docs/source/get-started/quickstart.mdx
@@ -10,9 +10,8 @@ You need:
- `Cognitive Services Contributor` access to create the deployment
- Managed compute quota for the accelerator that your model requires
-
- Managed compute runs on dedicated GPU capacity and bills for as long as the deployment exists. Delete the deployments that you no longer need.
-
+> [!WARNING]
+> Managed compute runs on dedicated GPU capacity and bills for as long as the deployment exists. Delete the deployments that you no longer need.
## Deploy a model
@@ -21,7 +20,7 @@ Pick the Foundry portal for a first, guided deployment, or Python for a repeatab
-1. Open [Microsoft Foundry](https://ai.azure.com) and select your project.
+1. Open Microsoft Foundry and select your project.
2. Go to **Discover → Models**, and filter **Collection** by **Hugging Face** and **Deployment options** by **Managed compute**.
3. Select a model and choose **Deploy**.
4. Enter a deployment name, pick a deployment template and its accelerator, and select **Deploy**.
@@ -77,9 +76,8 @@ The call returns when provisioning finishes, and prints `Succeeded`.
-
- The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
-
+> [!TIP]
+> The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
## Call the deployment
@@ -117,6 +115,6 @@ client.managed_compute_deployments.begin_delete(
## Next steps
-- Browse the [catalog](../catalog/qwen) to pick a model.
+- Browse the catalog to pick a model: [Qwen](../catalog/qwen), [DeepSeek AI](../catalog/deepseek-ai), [NVIDIA](../catalog/nvidia), [Microsoft](../catalog/microsoft).
- Turn reasoning on or off per request in [Enable and disable reasoning](../advanced/reasoning).
- Connect the deployment to a harness such as [Codex](../harness/codex) or [Hermes Agent](../harness/hermes-agent).
diff --git a/docs/source/harness/codex.mdx b/docs/source/harness/codex.mdx
index 9edd34b..27fb0eb 100644
--- a/docs/source/harness/codex.mdx
+++ b/docs/source/harness/codex.mdx
@@ -1,6 +1,6 @@
# Codex
-[Codex](https://developers.openai.com/codex/cli) is OpenAI's coding agent. The CLI works from your terminal, where it can inspect a repository, edit files, and run commands with your approval.
+Codex is OpenAI's coding agent. The CLI works from your terminal, where it can inspect a repository, edit files, and run commands with your approval.
@@ -18,17 +18,16 @@ On Windows, run:
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"
```
-You can also install Codex with [npm or Homebrew](https://github.com/openai/codex#installing-and-running-codex-cli).
+You can also install Codex with npm or Homebrew.
## Connect the model
-
- This guide assumes that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
-
+> [!TIP]
+> This guide assumes that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
-Codex allows you to [define custom profiles to connect to custom model providers](https://learn.chatgpt.com/docs/config-file/config-advanced) through the Responses API.
+Codex allows you to define custom profiles to connect to custom model providers through the Responses API.
-For example, suppose you [deploy Qwen3.8 27B on Microsoft Foundry](https://ai.azure.com/nextgen/goto/discover/models/qwen--qwen3.8-27b/version/2/registry/azure-huggingface) in a project named `huggingface-demo`, you can create a custom profile at `~/.codex/microsoft-foundry.config.toml` with:
+For example, suppose you deploy Qwen3.8 27B on Microsoft Foundry in a project named `huggingface-demo`, you can create a custom profile at `~/.codex/microsoft-foundry.config.toml` with:
```toml
model = "qwen--qwen3-8-27b"
@@ -56,7 +55,7 @@ For PowerShell, use:
$env:MICROSOFT_FOUNDRY_API_KEY=""
```
-You can copy the key from [Microsoft Foundry](https://ai.azure.com), or retrieve it with the Azure CLI:
+You can copy the key from Microsoft Foundry, or retrieve it with the Azure CLI:
```bash
az cognitiveservices account keys list \
diff --git a/docs/source/harness/hermes-agent.mdx b/docs/source/harness/hermes-agent.mdx
index 875b2aa..3fca9e5 100644
--- a/docs/source/harness/hermes-agent.mdx
+++ b/docs/source/harness/hermes-agent.mdx
@@ -1,12 +1,12 @@
# Hermes Agent
-[Hermes Agent](https://hermes-agent.nousresearch.com/) is a self-improving AI agent built by Nous Research. It's the only agent with a built-in learning loop, it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.
+Hermes Agent is a self-improving AI agent built by Nous Research. It's the only agent with a built-in learning loop, it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.
## Install the harness
-You can either install the desktop application for macOS from https://hermes-agent.nousresearch.com/ or rather install it programmatically for Linux or macOS as:
+You can either install the desktop application for macOS from https://hermes-agent.nousresearch.com/ or rather install it programmatically for Linux or macOS as:
```bash
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
@@ -20,15 +20,14 @@ irm https://hermes-agent.nousresearch.com/install.ps1 | iex
## Connect the model
-
- This guide assumes that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
-
+> [!TIP]
+> This guide assumes that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
Once installed, you can connect your OpenAI-compatible models deployed on Microsoft Foundry to it as it supports both Chat Completions API and Responses API (recommended) specifications.
-You need to create or add the following into your Hermes Agent configuration file at `~/.hermes/config.yaml`. More information at [Hermes Agent - Configuring Models](https://hermes-agent.nousresearch.com/docs/user-guide/configuring-models).
+You need to create or add the following into your Hermes Agent configuration file at `~/.hermes/config.yaml`. More information at Hermes Agent - Configuring Models.
-Let's assume that you want to [deploy Qwen3.8 27B on Microsoft Foundry](https://ai.azure.com/nextgen/goto/discover/models/qwen--qwen3.8-27b/version/2/registry/azure-huggingface) in a project named `huggingface-demo` then the `~/.hermes/config.yaml` would look like:
+Let's assume that you want to deploy Qwen3.8 27B on Microsoft Foundry in a project named `huggingface-demo` then the `~/.hermes/config.yaml` would look like:
```yaml
model:
@@ -54,9 +53,8 @@ agent:
qwen--qwen3-8-27b: medium
```
-
- As a reminder, the value for `HERMES_CUSTOM_MICROSOFT_FOUNDRY_API_KEY` is either obtained from [Microsoft Foundry](https://ai.azure.com) or programmatically with the `az` CLI as `az cognitiveservices account keys list --subscription --resource-group --name huggingface-demo -o tsv --query key1`.
-
+> [!TIP]
+> As a reminder, the value for `HERMES_CUSTOM_MICROSOFT_FOUNDRY_API_KEY` is either obtained from Microsoft Foundry or programmatically with the `az` CLI as `az cognitiveservices account keys list --subscription --resource-group --name huggingface-demo -o tsv --query key1`.
Alternatively, from the desktop application you can also open the settings and click on "Providers -> Custom Endpoints" and create the endpoint with the deployed model, then set the following:
diff --git a/docs/source/harness/index.mdx b/docs/source/harness/index.mdx
index dd2b217..ab2a205 100644
--- a/docs/source/harness/index.mdx
+++ b/docs/source/harness/index.mdx
@@ -2,67 +2,66 @@
An agent harness is the application that drives a model: it holds the conversation, exposes tools such as file edits or shell commands, and decides what to send back to the model on every turn. Any harness that speaks the OpenAI Chat Completions API or the Responses API can point at a model deployed on Microsoft Foundry.
-
- These guides assume that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
-
+> [!TIP]
+> These guides assume that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
Pick a harness to connect it to your Foundry deployment:
-
+
-
- Codex
- OpenAI's coding agent for the terminal
+
+ Codex
+ OpenAI's coding agent for the terminal
-
- Hermes Agent
- Nous Research's self-improving agent
+
+ Hermes Agent
+ Nous Research's self-improving agent
-
- Claude Code
- Coming soon
+
+ Claude Code
+ Coming soon
-
- Pi
- Coming soon
+
+ Pi
+ Coming soon
diff --git a/docs/source/index.mdx b/docs/source/index.mdx
index acd2a71..9252409 100644
--- a/docs/source/index.mdx
+++ b/docs/source/index.mdx
@@ -10,7 +10,7 @@ The collection lives in the `azure-huggingface` registry and shows up in the Fou
-Each model is published as a versioned asset in the registry, with one or more deployment templates that pin the serving engine, the context window, and the accelerator it runs on. Browse the collection by family in the [catalog](./catalog/qwen), where every entry links to the latest version of the model on Microsoft Foundry.
+Each model is published as a versioned asset in the registry, with one or more deployment templates that pin the serving engine, the context window, and the accelerator it runs on. Browse the collection by family in the catalog — [Qwen](./catalog/qwen), [DeepSeek AI](./catalog/deepseek-ai), [NVIDIA](./catalog/nvidia), and [Microsoft](./catalog/microsoft) — where every entry links to the latest version of the model on Microsoft Foundry.
## Deploy a model
diff --git a/docs/source/large-language-models/deepseek-v4-flash-0731.mdx b/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
index 9ccf596..9e4ad59 100644
--- a/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
+++ b/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
@@ -1,6 +1,6 @@
# DeepSeek V4 Flash 0731
-[`deepseek-ai/DeepSeek-V4-Flash-0731`](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731) is the official DeepSeek V4 Flash release, superseding the preview with stronger coding, tool-use, and agentic capabilities.
+
- The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
-
+> [!TIP]
+> The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
## Inference
@@ -127,4 +126,4 @@ Use `low` for quick tasks, `high` for most agentic work, and `max` for the harde
Connect the deployment to a compatible harness, such as [Hermes Agent](../harness/hermes-agent). Confirm that the harness supports Chat Completions and the model's reasoning controls.
-For more model and serving details, see the [DeepSeek model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731) and the [vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4-Flash).
+For more model and serving details, see the DeepSeek model card and the vLLM recipe.
diff --git a/docs/source/large-language-models/glm-5-2-fp8.mdx b/docs/source/large-language-models/glm-5-2-fp8.mdx
index 9f14a24..6bc8947 100644
--- a/docs/source/large-language-models/glm-5-2-fp8.mdx
+++ b/docs/source/large-language-models/glm-5-2-fp8.mdx
@@ -1,6 +1,6 @@
# GLM 5.2 FP8
-[`zai-org/GLM-5.2-FP8`](https://huggingface.co/zai-org/GLM-5.2-FP8) is an open-weight reasoning model for coding, tool use, and long-running agent tasks.
+
- The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
-
+> [!TIP]
+> The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.
## Inference
From 218cb1b144d9129706c5021b98ed926fdf950c18 Mon Sep 17 00:00:00 2001
From: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
Date: Wed, 19 Aug 2026 14:29:33 +0200
Subject: [PATCH 15/17] Update `_toctree.yml` and `harness/index.mdx`
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/_toctree.yml | 49 ++++++++++++++++++-----------------
docs/source/harness/index.mdx | 2 --
2 files changed, 25 insertions(+), 26 deletions(-)
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
index 960b3e9..efab873 100644
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -1,17 +1,16 @@
-- sections:
- - local: index
- title: Overview
+- local: index
title: Overview
-- sections:
- - local: catalog/qwen
- title: Qwen
- - local: catalog/deepseek-ai
- title: DeepSeek AI
- - local: catalog/nvidia
- title: NVIDIA
- - local: catalog/microsoft
- title: Microsoft
+- local: catalog/index
title: Catalog
+ sections:
+ - local: catalog/qwen
+ title: Qwen
+ - local: catalog/deepseek-ai
+ title: DeepSeek AI
+ - local: catalog/nvidia
+ title: NVIDIA
+ - local: catalog/microsoft
+ title: Microsoft
- sections:
- local: get-started/quickstart
title: Quickstart
@@ -22,21 +21,23 @@
- local: large-language-models/deepseek-v4-flash-0731
title: DeepSeek V4 Flash 0731
- local: large-language-models/qwen3-8-27b
- title: Qwen3.8 27B
+ title: Qwen3.8 27B (Coming soon)
title: Large Language Models
-- sections:
- - local: harness/index
- title: Overview
- - local: harness/claude-code
- title: Claude Code (Coming soon)
- - local: harness/codex
- title: Codex
- - local: harness/hermes-agent
- title: Hermes Agent
- - local: harness/pi
- title: Pi (Coming soon)
+- local: harness/index
title: Harness
+ sections:
+ - local: harness/index
+ title: Overview
+ - local: harness/claude-code
+ title: Claude Code (Coming soon)
+ - local: harness/codex
+ title: Codex
+ - local: harness/hermes-agent
+ title: Hermes Agent
+ - local: harness/pi
+ title: Pi (Coming soon)
- sections:
- local: advanced/reasoning
title: Enable and disable reasoning
title: Advanced
+ isExpanded: true
diff --git a/docs/source/harness/index.mdx b/docs/source/harness/index.mdx
index ab2a205..4422709 100644
--- a/docs/source/harness/index.mdx
+++ b/docs/source/harness/index.mdx
@@ -5,8 +5,6 @@ An agent harness is the application that drives a model: it holds the conversati
> [!TIP]
> These guides assume that you already have a model running on Microsoft Foundry. If you do not, follow the [quickstart](../get-started/quickstart) and deploy one from the Foundry portal or with Python first.
-Pick a harness to connect it to your Foundry deployment:
-
-
- Codex
- OpenAI's coding agent for the terminal
+
+ Codex
+ OpenAI's coding agent for the terminal
-
- Hermes Agent
- Nous Research's self-improving agent
+
+ Hermes Agent
+ Nous Research's self-improving agent
-
- Claude Code
- Coming soon
+
+ Claude Code
+ Coming soon
-
- Pi
- Coming soon
+
+ Pi
+ Coming soon
deepseek-ai/DeepSeek-V4-Flash-0731 is the official DeepSeek V4 Flash release, superseding the preview with stronger coding, tool-use, and agentic capabilities.
DeepSeek V4 Flash is a mixture-of-experts model with 284B total parameters and 13B activated per token. It combines DeepSeek's compressed attention architecture with a fused DSpark speculative-decoding module, keeping inference efficient while supporting long-context workloads. The model is released under the MIT license.
@@ -72,9 +72,8 @@ deployment = client.managed_compute_deployments.begin_create_or_update(
print(deployment.properties.provisioning_state)
```
-zai-org/GLM-5.2-FP8 is an open-weight reasoning model for coding, tool use, and long-running agent tasks.
@@ -61,9 +61,8 @@ deployment = client.managed_compute_deployments.begin_create_or_update(
print(deployment.properties.provisioning_state)
```
-
Date: Wed, 19 Aug 2026 16:18:31 +0200
Subject: [PATCH 16/17] Add catalog index and move examples
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/_toctree.yml | 57 ++++++------
docs/source/catalog/index.mdx | 88 +++++++++++++++++++
.../deepseek-v4-flash-0731.mdx | 2 +-
.../glm-5-2-fp8.mdx | 2 +-
docs/source/examples/qwen3-8-27b.mdx | 3 +
docs/source/get-started/quickstart.mdx | 2 +-
docs/source/harness/index.mdx | 40 ++++-----
docs/source/index.mdx | 2 +-
.../large-language-models/qwen3-8-27b.mdx | 3 -
9 files changed, 146 insertions(+), 53 deletions(-)
create mode 100644 docs/source/catalog/index.mdx
rename docs/source/{large-language-models => examples}/deepseek-v4-flash-0731.mdx (98%)
rename docs/source/{large-language-models => examples}/glm-5-2-fp8.mdx (98%)
create mode 100644 docs/source/examples/qwen3-8-27b.mdx
delete mode 100644 docs/source/large-language-models/qwen3-8-27b.mdx
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
index efab873..d257e84 100644
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -1,43 +1,48 @@
- local: index
title: Overview
+
- local: catalog/index
title: Catalog
+ isExpanded: true
sections:
- - local: catalog/qwen
- title: Qwen
- - local: catalog/deepseek-ai
- title: DeepSeek AI
- - local: catalog/nvidia
- title: NVIDIA
- - local: catalog/microsoft
- title: Microsoft
+ - local: catalog/qwen
+ title: Qwen
+ - local: catalog/deepseek-ai
+ title: DeepSeek AI
+ - local: catalog/nvidia
+ title: NVIDIA
+ - local: catalog/microsoft
+ title: Microsoft
+
- sections:
- local: get-started/quickstart
title: Quickstart
title: Get started
+
- sections:
- - local: large-language-models/glm-5-2-fp8
- title: GLM 5.2 FP8
- - local: large-language-models/deepseek-v4-flash-0731
- title: DeepSeek V4 Flash 0731
- - local: large-language-models/qwen3-8-27b
- title: Qwen3.8 27B (Coming soon)
- title: Large Language Models
+ - local: examples/glm-5-2-fp8
+ title: Open frontier-scale intelligence with GLM 5.2 and vLLM on Foundry
+ - local: examples/deepseek-v4-flash-0731
+ title: Long-context reasoning and agentic coding with DeepSeek V4 Flash on Foundry
+ - local: examples/qwen3-8-27b
+ title: Multimodal understanding with Qwen3.8 27B on Foundry
+ title: Examples
+
- local: harness/index
title: Harness
+ isExpanded: true
sections:
- - local: harness/index
- title: Overview
- - local: harness/claude-code
- title: Claude Code (Coming soon)
- - local: harness/codex
- title: Codex
- - local: harness/hermes-agent
- title: Hermes Agent
- - local: harness/pi
- title: Pi (Coming soon)
+ - local: harness/claude-code
+ title: Claude Code (Coming soon)
+ - local: harness/codex
+ title: Codex
+ - local: harness/hermes-agent
+ title: Hermes Agent
+ - local: harness/pi
+ title: Pi (Coming soon)
+
- sections:
- local: advanced/reasoning
title: Enable and disable reasoning
title: Advanced
- isExpanded: true
+ isExpanded: false
diff --git a/docs/source/catalog/index.mdx b/docs/source/catalog/index.mdx
new file mode 100644
index 0000000..74ff012
--- /dev/null
+++ b/docs/source/catalog/index.mdx
@@ -0,0 +1,88 @@
+# Catalog
+
+Hugging Face publishes a curated collection of open models to the `azure-huggingface` registry on Microsoft Foundry. Each one is packaged with the serving engine, context window, and accelerator it was validated on, so you deploy it on managed compute without assembling the stack yourself.
+
+The families below have a page here, listing every one of their models available on Foundry:
+
+
-
- Codex
- OpenAI's coding agent for the terminal
+
+ Codex
+ OpenAI's coding agent for the terminal
-
- Hermes Agent
- Nous Research's self-improving agent
+
+ Hermes Agent
+ Nous Research's self-improving agent
-
- Claude Code
- Coming soon
+
+ Claude Code
+ Coming soon
-
- Pi
- Coming soon
+
+ Pi
+ Coming soon
diff --git a/docs/source/index.mdx b/docs/source/index.mdx
index 9252409..998faf2 100644
--- a/docs/source/index.mdx
+++ b/docs/source/index.mdx
@@ -10,7 +10,7 @@ The collection lives in the `azure-huggingface` registry and shows up in the Fou
-Each model is published as a versioned asset in the registry, with one or more deployment templates that pin the serving engine, the context window, and the accelerator it runs on. Browse the collection by family in the catalog — [Qwen](./catalog/qwen), [DeepSeek AI](./catalog/deepseek-ai), [NVIDIA](./catalog/nvidia), and [Microsoft](./catalog/microsoft) — where every entry links to the latest version of the model on Microsoft Foundry.
+Each model is published as a versioned asset in the registry, with one or more deployment templates that pin the serving engine, the context window, and the accelerator it runs on. Browse it in the [catalog](./catalog/index), where every entry links to the latest version of the model on Microsoft Foundry.
## Deploy a model
diff --git a/docs/source/large-language-models/qwen3-8-27b.mdx b/docs/source/large-language-models/qwen3-8-27b.mdx
deleted file mode 100644
index a6cabf2..0000000
--- a/docs/source/large-language-models/qwen3-8-27b.mdx
+++ /dev/null
@@ -1,3 +0,0 @@
-# Qwen3.8 27B
-
-Coming soon!
From 8148c54a55da34bc6d62066f77204fdeac1d2d96 Mon Sep 17 00:00:00 2001
From: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
Date: Wed, 19 Aug 2026 16:27:49 +0200
Subject: [PATCH 17/17] Update examples and put `get-stated` before catalog
Signed-off-by: Alvaro Bartolome <36760800+alvarobartt@users.noreply.github.com>
---
docs/source/_toctree.yml | 10 ++++----
.../examples/deepseek-v4-flash-0731.mdx | 25 ++++++++++++-------
docs/source/examples/glm-5-2-fp8.mdx | 2 +-
3 files changed, 22 insertions(+), 15 deletions(-)
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
index d257e84..30251fd 100644
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -1,6 +1,11 @@
- local: index
title: Overview
+- sections:
+ - local: get-started/quickstart
+ title: Quickstart
+ title: Get started
+
- local: catalog/index
title: Catalog
isExpanded: true
@@ -14,11 +19,6 @@
- local: catalog/microsoft
title: Microsoft
-- sections:
- - local: get-started/quickstart
- title: Quickstart
- title: Get started
-
- sections:
- local: examples/glm-5-2-fp8
title: Open frontier-scale intelligence with GLM 5.2 and vLLM on Foundry
diff --git a/docs/source/examples/deepseek-v4-flash-0731.mdx b/docs/source/examples/deepseek-v4-flash-0731.mdx
index 34c01e5..2d5c749 100644
--- a/docs/source/examples/deepseek-v4-flash-0731.mdx
+++ b/docs/source/examples/deepseek-v4-flash-0731.mdx
@@ -1,17 +1,24 @@
-# Long-context reasoning and agentic coding with DeepSeek V4 Flash on Foundry
+# Long-context reasoning and agentic coding with DeepSeek V4 Flash
+
+
+
+ Qwen
+ Chat, vision-language, coding, embeddings, and image generation
+
+
+
+ 
+
+ DeepSeek AI
+ Mixture-of-experts models for long-context reasoning and coding
+
+
+
+
+
+ NVIDIA
+ Nemotron and Cosmos, plus FP8 and NVFP4 quantizations
+
+
+
+
+
+ Microsoft
+ Phi small language models and domain-tuned variants
+
+
+
+
+## And much more
+
+Those pages cover 50 of the 102 models in the collection. The rest spans many more providers, including:
+
+- **Z.ai** GLM 5.2, and **OpenAI** gpt-oss 20B and 120B, for open frontier-scale chat and reasoning
+- **Google** Gemma 4, **Mistral AI**, **IBM Granite**, and **Cohere Labs** North, for general-purpose chat
+- **OpenAI** Whisper for speech recognition, and **Qwen** and **Tongyi-MAI** models for image generation
+- **Sentence Transformers** all-MiniLM, **Answer.AI** ModernBERT, and **BERT** variants, for embeddings and classification at a fraction of the cost of a chat model
+
+New models land in the collection regularly, so the Foundry catalog is always the complete, current list:
+
+
+
+ Browse the full collection on Microsoft Foundry
+ Filter by Collection: Hugging Face and Deployment options: Managed compute to see all 102
+
+
+
diff --git a/docs/source/large-language-models/deepseek-v4-flash-0731.mdx b/docs/source/examples/deepseek-v4-flash-0731.mdx
similarity index 98%
rename from docs/source/large-language-models/deepseek-v4-flash-0731.mdx
rename to docs/source/examples/deepseek-v4-flash-0731.mdx
index 9e4ad59..34c01e5 100644
--- a/docs/source/large-language-models/deepseek-v4-flash-0731.mdx
+++ b/docs/source/examples/deepseek-v4-flash-0731.mdx
@@ -1,4 +1,4 @@
-# DeepSeek V4 Flash 0731
+# Long-context reasoning and agentic coding with DeepSeek V4 Flash on Foundry
+
+ Qwen
+ Chat, vision-language, coding, embeddings, and image generation
+
+
+
+
+
+ DeepSeek AI
+ Mixture-of-experts models for long-context reasoning and coding
+
+
+
+
+
+ NVIDIA
+ Nemotron and Cosmos, plus FP8 and NVFP4 quantizations
+
+
+
+
+
+ Microsoft
+ Phi small language models and domain-tuned variants
+
+
+deepseek-ai/DeepSeek-V4-Flash-0731 is the official DeepSeek V4 Flash release, superseding the preview with stronger coding, tool-use, and agentic capabilities.
diff --git a/docs/source/large-language-models/glm-5-2-fp8.mdx b/docs/source/examples/glm-5-2-fp8.mdx
similarity index 98%
rename from docs/source/large-language-models/glm-5-2-fp8.mdx
rename to docs/source/examples/glm-5-2-fp8.mdx
index 6bc8947..a2f6917 100644
--- a/docs/source/large-language-models/glm-5-2-fp8.mdx
+++ b/docs/source/examples/glm-5-2-fp8.mdx
@@ -1,4 +1,4 @@
-# GLM 5.2 FP8
+# Open frontier-scale intelligence with GLM 5.2 and vLLM on Foundry
zai-org/GLM-5.2-FP8 is an open-weight reasoning model for coding, tool use, and long-running agent tasks.
diff --git a/docs/source/examples/qwen3-8-27b.mdx b/docs/source/examples/qwen3-8-27b.mdx
new file mode 100644
index 0000000..18ecea1
--- /dev/null
+++ b/docs/source/examples/qwen3-8-27b.mdx
@@ -0,0 +1,3 @@
+# Multimodal understanding with Qwen3.8 27B on Foundry
+
+Coming soon!
diff --git a/docs/source/get-started/quickstart.mdx b/docs/source/get-started/quickstart.mdx
index 52b0f7f..2237e17 100644
--- a/docs/source/get-started/quickstart.mdx
+++ b/docs/source/get-started/quickstart.mdx
@@ -115,6 +115,6 @@ client.managed_compute_deployments.begin_delete(
## Next steps
-- Browse the catalog to pick a model: [Qwen](../catalog/qwen), [DeepSeek AI](../catalog/deepseek-ai), [NVIDIA](../catalog/nvidia), [Microsoft](../catalog/microsoft).
+- Browse the [catalog](../catalog/index) to pick a model.
- Turn reasoning on or off per request in [Enable and disable reasoning](../advanced/reasoning).
- Connect the deployment to a harness such as [Codex](../harness/codex) or [Hermes Agent](../harness/hermes-agent).
diff --git a/docs/source/harness/index.mdx b/docs/source/harness/index.mdx
index 4422709..da8d1b0 100644
--- a/docs/source/harness/index.mdx
+++ b/docs/source/harness/index.mdx
@@ -8,58 +8,58 @@ An agent harness is the application that drives a model: it holds the conversati
-
- Codex
- OpenAI's coding agent for the terminal
+
+ Codex
+ OpenAI's coding agent for the terminal
-
- Hermes Agent
- Nous Research's self-improving agent
+
+ Hermes Agent
+ Nous Research's self-improving agent
-
- Claude Code
- Coming soon
+
+ Claude Code
+ Coming soon
-
- Pi
- Coming soon
+
+ Pi
+ Coming soon
deepseek-ai/DeepSeek-V4-Flash-0731 is the official DeepSeek V4 Flash release, superseding the preview with stronger coding, tool-use, and agentic capabilities.
DeepSeek V4 Flash is a mixture-of-experts model with 284B total parameters and 13B activated per token. It combines DeepSeek's compressed attention architecture with a fused DSpark speculative-decoding module, keeping inference efficient while supporting long-context workloads. The model is released under the MIT license.
-DeepSeek reports the following results for the `0731` release:
-
-| Benchmark | Score |
-|---|---:|
-| Terminal Bench 2.1 | 82.7 |
-| DeepSWE | 54.4 |
-| Toolathlon-Verified | 70.3 |
-| Agents' Last Exam | 25.2 |
+
+
+| Benchmark | DeepSeek-V4-Flash-0731 | DeepSeek-V4-Flash (Preview) | DeepSeek-V4-Pro (Preview) | GLM-5.2 | Opus-4.8 |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 | 81.0 | 85.0 |
+| NL2Repo | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 |
+| Cybergym | 76.7 | 38.7 | 52.7 | - | 83.1 |
+| DeepSWE | 54.4 | 7.3 | 12.8 | 46.2 | 58.0 |
+| Toolathlon-Verified | 70.3 | 49.7 | 55.9 | 59.9 | 76.2 |
+| Agents' Last Exam | 25.2 | 15.8 | 16.5 | 23.8 | 25.7 |
+| AutomationBench Public | 25.1 | 10.8 | 12.8 | 12.9 | 27.2 |
+| DSBench-FullStack † | 68.7 | 37.0 | 41.8 | 61.8 | 71.6 |
+| DSBench-Hard † | 59.6 | 25.8 | 31.1 | 54.5 | 71.7 |
+
+
On Microsoft Foundry, it runs with vLLM using:
diff --git a/docs/source/examples/glm-5-2-fp8.mdx b/docs/source/examples/glm-5-2-fp8.mdx
index a2f6917..233bdfb 100644
--- a/docs/source/examples/glm-5-2-fp8.mdx
+++ b/docs/source/examples/glm-5-2-fp8.mdx
@@ -1,4 +1,4 @@
-# Open frontier-scale intelligence with GLM 5.2 and vLLM on Foundry
+# Open frontier-scale intelligence with GLM 5.2
zai-org/GLM-5.2-FP8 is an open-weight reasoning model for coding, tool use, and long-running agent tasks.