Skip to content

Available models: Azure Local Foundry

One of three availability references

This page covers Azure Local Foundry, cluster-scale inference on Azure Local. The other two targets carry different rosters: Azure AI Foundry (hosted cloud) and Foundry Local (a runtime inside your own application). Compare all three on Deployment targets.

Everything here is drawn from SPIKE-22, which read the published catalog snapshot dated 2026-07-28, with GPU figures confirmed against Microsoft's published vLLM reference.

Nothing here is deployed

This target has no deployment behind it in this repository, so no row on this page is deployed. Everything in the catalog and deployable in principle is available. This is an availability reference, not an as-built record.

What Azure Local Foundry is, precisely

An Azure Arc-enabled Kubernetes extension, not a device runtime. An operator control plane watches cluster state and reconciles model resources through declarative custom resources (Model and ModelDeployment CRDs).

DeploymentArc extension, or Helm
AvailabilityPreview, by request. Submit the access request form before planning around it
RuntimesONNX-GenAI (CPU or GPU) and vLLM (GPU only, high throughput)
APIOpenAI-compatible: /v1/chat/completions generative, /v1/predict predictive
Azure subscriptionRequired, plus Entra permissions and a supported region
AuthAPI keys or Microsoft Entra ID, TLS-terminated gateway
DisconnectedSupported, with a deployment model consistent with connected scenarios

It also supports bring-your-own models containerized with ONNX or vLLM, including Hugging Face models and fine-tuned variants. Predictive workloads support BYO models only; catalog models cannot serve predictive tasks.

What this target does not have

Text, reasoning, code, speech-to-text, and vision input. Confirmed absent (SPIKE-22 Q5):

CapabilityAvailable on Azure Local Foundry
Image generationNo
Text to speechNo
EmbeddingsNo
Video generationNo
Vision inputYes - and this is exclusive to this target

Nothing in this project's cloud catalog runs here.

Two notes before you read the tables

Sizes and licences are mostly UNKNOWN, deliberately. The published catalog snapshot carries no size column, and licence values are published for exactly one model. Resolving either is a query against a running cluster, not a research question. This repository marks unverifiable facts UNKNOWN rather than inventing them.

Publisher reads Microsoft on every ONNX row. That field means "publisher of the ONNX build," not the originating lab.

What each family is actually for

The alias tells you the size, the vendor, and often the quantization. It does not tell you what the model is good at. This section does.

FamilyReach for it whenAvoid it when
Phi (Microsoft)You want the best quality per gigabyte of GPU. phi-4-mini-instruct fits in 7.8 GB with a 93,520 context. phi-4-reasoning is cluster-exclusive and the choice for multi-step working.You need broad world knowledge or very long documents.
gpt-oss-20b / 120b (OpenAI, open weight)Strong function calling, structured output, and reasoning. gpt-oss-20b is Microsoft's recommended model for Agentic Retrieval. 120b when quality outranks cost.Your GPUs are Ampere. See the GPU warning below, and note Microsoft says gpt-oss-20b "requires its own GPU".
DeepSeek R1 distillsReasoning-heavy analysis where you want visible working, at 1.5B/7B/14B so you can size to the node.Latency matters: reasoning models emit far more tokens before answering.
DeepSeek V3 lineFrontier-class open-weight quality on-premises. Cluster-exclusive.Your cluster is small. These are very large and will dominate a node's memory.
Qwen (Alibaba)Multilingual work, and precise size selection from 0.6B to 32B. qwen3-32b is cluster-exclusive.Licence certainty matters and you have not checked.
Qwen CoderCode generation and developer tooling, 0.5B to 14B.General chat: these are worse at prose than same-size siblings.
Mistral denseGeneral chat with widely-adopted open models, scaling from 7B to mistral-large-3-675b, the largest entry in the catalog.Memory is tight. mistral-7b-v0.2 alone needs 15.64 GB, roughly double Phi-4-mini.
Mistral MoE and specialistMixture-of-experts throughput (mixtral), agentic coding (devstral), maths (mathstral), reasoning (magistral, ministral-3-14b-reasoning).You have left vllm.modelCacheStorageGi at its 100 GiB default: Microsoft names magistral as a model that exceeds it.
Pixtral, Nemotron VLVision input. Document understanding, screenshots, diagrams, image Q and A. These are the only vision-capable models on any on-premises target.You only need text. Vision models cost more memory for no benefit.
Nemotron general and omni (NVIDIA)High-throughput GPU serving, tuned for concurrency rather than single-request latency. Quantization is in the alias: bf16 quality, fp8 balance, nvfp4 smallest.You are on non-NVIDIA hardware or older generations.
Nemotron code / math / reasoningNarrow, deep specialisms: competitive coding (opencodereasoning, -ioi), terminal and tool use (nemotron-terminal), mathematics (openmath, acemath), reasoning (openreasoning, acereason).You want one model for everything. These are deliberately narrow.
OLMo, SmolLMFully-open provenance (OLMo, Allen Institute; olmo-3.1-32b is cluster-exclusive) or the smallest viable footprint (SmolLM3, 3B).You need frontier quality.
Whisper (ASR)Speech to text. -tiny and -base for throughput, -large-v3-turbo for accuracy.You want text to speech. Nothing on this target does TTS.
Nemotron streaming ASRLive transcription where Whisper's batch shape adds latency.You need languages beyond English.

Start here

phi-4-mini-instruct for general use, qwen2.5-coder-7b for code, gpt-oss-20b if you are building Agentic Retrieval and have the GPU for it, pixtral-12b-2409 if you need to read images, whisper-large-v3-turbo for transcription.

Section 1: the ONNX roster

35 aliases, 70 catalog entries, served by the ONNX-GenAI runtime on CPU or GPU. Every model in this table also runs on Foundry Local; this is the shared core the two on-premises targets have in common.

Model (alias)Publisher / originRuntimeNotesStatus
phi-4-miniMicrosoft / Microsoftonnx-genai, cpu or gpuAlso a vLLM entry; MIT licence; 7.806 GB GPU memory on vLLMavailable
phi-4-mini-reasoningMicrosoft / Microsoftonnx-genaiAlso a vLLM entry; 7.806 GB GPU memoryavailable
phi-4Microsoft / Microsoftonnx-genaiAlso a vLLM entryavailable
phi-3.5-miniMicrosoft / Microsoftonnx-genaiAlso a vLLM entry; 8.428 GB GPU memoryavailable
phi-3-mini-4kMicrosoft / Microsoftonnx-genai3.8B paramsavailable
phi-3-mini-128kMicrosoft / Microsoftonnx-genai3.8B paramsavailable
gpt-oss-20bMicrosoft / OpenAI (open weight)onnx-genaiAlso a vLLM entry; see the GPU warning belowavailable
qwen3-0.6b / -1.7b / -4b / -8b / -14bMicrosoft / Alibabaonnx-genaiqwen3-4b has CPU/GPU version skewavailable
qwen3.5-2b-textMicrosoft / Alibabaonnx-genaiavailable
qwen2.5-0.5b / -1.5b / -7b / -14bMicrosoft / Alibabaonnx-genaiAlso vLLM entriesavailable
qwen2.5-coder-0.5b / -1.5b / -7b / -14bMicrosoft / Alibabaonnx-genaiAlso vLLM entriesavailable
deepseek-r1-7b / -14bMicrosoft / DeepSeekonnx-genaiDistilled Qwen; also vLLM entriesavailable
mistral-7b-v0.2Microsoft / Mistral AIonnx-genaiAlso a vLLM entry; 15.64 GB GPU memoryavailable
mistral-nemo-12b-instructMicrosoft / Mistral AI and NVIDIAonnx-genaiavailable
olmo-3-7b-instructMicrosoft / Allen Institute for AIonnx-genaiavailable
smollm3-3bMicrosoft / Hugging Faceonnx-genaiavailable
whisper-tiny / -base / -small / -medium / -large-v3-turboMicrosoft / OpenAIonnx-genaitask automatic-speech-recognitionavailable (ASR)
nemotron-speech-streaming-en-0.6b / -es-0.6bMicrosoft / NVIDIAonnx-genai0.6B paramsavailable (ASR)
nemotron-3.5-asr-streaming-0.6bMicrosoft / NVIDIAonnx-genai0.6B paramsavailable (ASR)

Section 2: the vLLM roster, exclusive to this target

100 entries, GPU only, and not available on Foundry Local at all: vLLM is a GPU-only container runtime with no device-SDK equivalent. Grouped by family, with every alias named. Framework is vllm and task is chat-completion except where noted. Size, RAM, and licence are UNKNOWN except where GPU memory is published.

FamilyAliasesNotesStatus
Microsoft Phiphi-4, phi-4-reasoning, phi-4-mini-instruct, phi-4-mini-reasoning, phi-3.5-mini-instructphi-4-reasoning has no ONNX entry, so it is exclusive to this targetavailable
OpenAI open weightgpt-oss-20b, gpt-oss-120bgpt-oss-120b is exclusive to this target. gpt-oss-20b is Microsoft's recommended model for Agentic Retrieval and "requires its own GPU"available
DeepSeekdeepseek-r1-distill-qwen-1.5b, -7b, -14b, deepseek-v3-0324, deepseek-v3.1, deepseek-v3.2, deepseek-v3.2-specialeThe V3 line is exclusive to this target and very largeavailable
Qwenqwen2.5-0.5b-instruct, -1.5b-instruct, -7b-instruct, -14b-instruct, qwen2.5-coder-0.5b/1.5b/7b/14b-instruct, qwen3-0.6b, -1.7b, -8b, -14b, -32bqwen3-32b is exclusive to this targetavailable
Mistral AI, densemistral-7b-instruct-v0.2, -v0.3, mistral-nemo-instruct-2407, mistral-nemo-instruct-fp8-2407, mistral-small-24b-instruct-2501, mistral-small-3.1-24b-instruct-2503, mistral-small-3.2-24b-instruct-2506, mistral-small-4-119b-2603, mistral-small-4-119b-2603-nvfp4, mistral-large-3-675b-instruct-2512mistral-large-3-675b is the largest entry in the catalogavailable
Mistral AI, MoE and specialistmixtral-8x7b-instruct-v0.1, mixtral-8x22b-instruct-v0.1, magistral-small-2506, -2507, -2509, devstral-small-2505, -2507, mathstral-7b-v0.1, ministral-3-3b-instruct-2512, ministral-3-8b-instruct-2512, ministral-3-14b-reasoning-2512Microsoft names magistral as a model that exceeds the 100 GiB vllm.modelCacheStorageGi defaultavailable
Mistral AI, visionpixtral-12b-2409Vision-languageavailable
NVIDIA Nemotron, visionnemotron-nano-12b-v2-vl-bf16, -fp8, -nvfp4-qadVision-languageavailable
NVIDIA Nemotron, omninemotron-3-nano-omni-30b-a3b-reasoning-bf16, -fp8, -nvfp4Multi-modal by name; capability surface not documentedavailable
NVIDIA Nemotron, generalnemotron-nano-9b-v2, -fp8, -nvfp4, -japanese, nemotron-nano-12b-v2, nemotron-3-nano-4b-bf16, -fp8, nemotron-3-nano-30b-a3b-bf16, -nvfp4, nemotron-3-super-120b-a12b-bf16, -fp8, -nvfp4, nemotron-4-mini-hindi-4b-instructQuantization is encoded in the aliasavailable
NVIDIA Nemotron, codeopencodereasoning-nemotron-7b, -14b, -32b, -32b-ioi, opencodereasoning-nemotron-1.1-7b, -1.1-14b, -1.1-32b, nemotron-terminal-8b, -14b, -32bavailable
NVIDIA Nemotron, math and reasoningopenmath-nemotron-1.5b, -7b, -14b, -14b-kaggle, -32b, openreasoning-nemotron-1.5b, -7b, -14b, -32b, acereason-nemotron-7b, -14b, acereason-nemotron-1.1-7b, acemath-rl-nemotron-7bavailable
Allen Institute for AIolmo-3-7b-instruct, olmo-3.1-32b-instructolmo-3.1-32b is exclusive to this targetavailable
Hugging Facesmollm3-3bavailable
OpenAI Whisper (ASR)whisper-tiny, whisper-base, whisper-small, whisper-medium, whisper-large-v3-turbotask automatic-speech-recognitionavailable (ASR)
NVIDIA streaming ASRnemotron-speech-streaming-en-0.6btask automatic-speech-recognitionavailable (ASR)

Vision input is exclusive to this target. pixtral-12b-2409 and the three nemotron-nano-12b-v2-vl variants are vision-language models and appear only here. Foundry Local has no vision-capable entry.

Published GPU requirements: size the cluster from these

These five are the only entries Microsoft publishes hard figures for. Read the GPU generation column before sizing anything.

ModelMax contextRecommended minimum GPURequired GPU memoryGPU utilization
Phi-3.5-mini-instruct29,472Ampere (CC 8.0)+8.428 GB0.85
Phi-4-mini-instruct93,520Ampere (CC 8.0)+7.806 GB0.85
Phi-4-mini-reasoning93,520Ampere (CC 8.0)+7.806 GB0.85
Mistral-7B-Instruct-v0.229,328Ampere (CC 8.0)+15.64 GB0.85
gpt-oss-20b96,784Blackwell (CC 10.0)+14.793 GB0.8

gpt-oss-20b is two GPU generations ahead of everything else

Every other documented model recommends Ampere (CC 8.0). gpt-oss-20b recommends Blackwell (CC 10.0), and it is also the model Microsoft recommends for Agentic Retrieval. Sizing a cluster from the other four rows and adding gpt-oss-20b later is an expensive surprise. Minimum supported across all five is Volta (CC 7.0), but minimum supported is not the same as usable.

The authoritative list is your cluster, not this page

No published Microsoft page is a complete catalog. concept-models shows five models and calls them "representative examples only"; concept-model-catalog shows twelve and says the table "isn't exhaustive." The two disagree in scope for the same product. Ask the cluster:

bash
kubectl get configmap foundry-local-catalog -n foundry-local-operator \
  -o jsonpath="{.data['catalog\.json']}" | ConvertFrom-Json |
  Select-Object -ExpandProperty models |
  Format-Table alias, displayName, task, framework

Or through the REST API, with port forwarding to inference-operator-api:

bash
curl -k -s https://localhost:8080/api/v1/models -H "Authorization: Bearer $token"

A capability read from documentation is a claim; exercised against the cluster it is a fact.

Keeping this current

This page is a transcription of a catalog snapshot dated 2026-07-28. It will drift, and the product is in preview under active deployment. Generating it from the live catalog is tracked as a feature request: see issue #15, auto-refresh the model availability catalogs.

See also