Appearance
Available models: Foundry Local
One of three availability references
This page covers Foundry Local, the runtime that ships inside your own application on a machine you own. The other two targets carry different rosters: Azure AI Foundry (hosted cloud) and Azure Local Foundry (cluster scale on Azure Local). Compare all three on Deployment targets.
Everything here is drawn from SPIKE-22, which read the published catalog snapshot dated 2026-07-28.
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. For what this project actually chose and runs, see the model catalog.
What Foundry Local is, precisely
It is a library that ships inside your application, not a server product. The runtime adds roughly 20 MB to your package and loads in-process through a C#, JavaScript, Python or Rust SDK. There is no separate installer or background service for end users to manage.
| Supported platforms | Windows 10 and later, and cross-platform: macOS and Linux |
| Install (CLI, for exploration) | winget install Microsoft.FoundryLocal, or brew tap microsoft/foundrylocal |
| Runtime | ONNX Runtime, with WinML on Windows for execution-provider registration |
| Model acquisition | Cloud-hosted Foundry Catalog, downloaded on first use, then cached and run fully offline |
| Azure subscription | Not required |
| Server edition | There is no Windows Server SKU. Microsoft scopes this as PCs and cross-platform devices. Running it on Windows Server is running the device runtime on server hardware, not a distinct product. For a server-class, cluster-scale deployment, the product is Azure Local Foundry. |
The optional local server is an OpenAI-compatible HTTP endpoint you can start inside your process for tools like LangChain or Open WebUI. It is not required for SDK use.
What this target does not have
Text, reasoning, code, and speech-to-text only. Confirmed absent (SPIKE-22 Q5):
| Capability | Available on Foundry Local |
|---|---|
| Image generation | No |
| Text to speech | No |
| Embeddings | No |
| Video generation | No |
| Vision input | No. Vision-language models are an Azure Local Foundry exclusive |
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 install, not a research question, and back-filling from upstream model cards would be guessing. 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. The originating lab is given after the slash where it differs.
What each family is actually for
The alias tells you the size and the vendor. It does not tell you what the model is good at. This section does.
| Family | Reach for it when | Avoid it when |
|---|---|---|
| Phi (Microsoft) | You want the best quality per gigabyte. phi-4-mini is the sensible default on a device: it is the only row with a confirmed MIT licence, and it fits in 3.6 GB on GPU. phi-4-mini-reasoning trades speed for multi-step working. | You need broad world knowledge or long multi-document context. Small models trade breadth for efficiency. |
| Qwen (Alibaba) | You need multilingual work, or you want to tune size precisely: the family spans 0.6B to 14B, so you can dial quality against RAM on the exact hardware you have. | Licence certainty matters and you have not checked. Licence is UNKNOWN on every Qwen row here. |
| Qwen Coder | Code generation, completion, refactoring, and developer tooling. qwen2.5-coder-7b is the usual balance point; -14b if the machine can hold it. | You want general chat. These are tuned for code and are worse at prose than their non-coder siblings of the same size. |
| DeepSeek R1 distills | Reasoning-heavy tasks where you want visible working: analysis, multi-step derivation, planning. Distilled onto Qwen, so they inherit its efficiency. | Latency matters. Reasoning models emit far more tokens before answering, and on a device that is wall-clock time. |
| Mistral | General-purpose chat with a widely-understood open model. mistral-7b-v0.2 is the most broadly adopted instruct model in the roster. | Memory is tight. It is the heaviest documented entry at 15.64 GB GPU memory, roughly double Phi-4-mini. |
| gpt-oss-20b | You want the largest open-weight model that still runs locally, with strong function calling and structured output. | Your hardware is modest. See the GPU note below. |
| OLMo, SmolLM | Fully-open-provenance work (OLMo, Allen Institute) or the smallest viable footprint (SmolLM3, 3B). Good for embedded and edge cases where 7B will not fit. | You need frontier quality. These are chosen for openness and size, not for benchmark scores. |
| Whisper (ASR) | Speech to text. whisper-tiny and -base for real-time on modest hardware; -large-v3-turbo when accuracy matters more than latency. | You want text to speech. Nothing on this target does TTS. |
| Nemotron streaming ASR (NVIDIA) | Live transcription, where Whisper's batch-oriented shape adds latency. English and Spanish variants at 0.6B. | You need languages beyond those published. |
Start here
phi-4-mini for general use, qwen2.5-coder-7b for code, whisper-base for transcription. All three run on modest hardware, and the first has a confirmed permissive licence.
gpt-oss-20b is not a modest-hardware model
On the cluster target its vLLM entry recommends Blackwell (CC 10.0) GPUs and 14.793 GB of GPU memory. The ONNX build here will run on less, but "runs" and "runs usefully" are different claims, and no device-side figure is published. Measure it on your hardware before designing around it.
Section 1: the ONNX roster
35 aliases, 70 catalog entries. Every row has both a -generic-cpu and a -cuda-gpu entry under the same base name. Every model in this table also runs on Azure Local Foundry; this is the shared core the two on-premises targets have in common.
| Model (alias) | Publisher / origin | Variant and execution provider | Size and RAM | Licence | Status |
|---|---|---|---|---|---|
phi-4-mini | Microsoft / Microsoft | Phi-4-mini-instruct-generic-cpu:5 CPU; -cuda-gpu:5 CUDA | 4.8 GB CPU, 3.6 GB GPU; RAM UNKNOWN | MIT | available |
phi-4-mini-reasoning | Microsoft / Microsoft | Phi-4-mini-reasoning-generic-cpu:3; -cuda-gpu:3 | UNKNOWN | UNKNOWN | available |
phi-4 | Microsoft / Microsoft | Phi-4-generic-cpu:2; -cuda-gpu:2 | UNKNOWN | UNKNOWN | available |
phi-3.5-mini | Microsoft / Microsoft | Phi-3.5-mini-instruct-generic-cpu:2; -cuda-gpu:2 | 2.53 GB device download | UNKNOWN | available |
phi-3-mini-4k | Microsoft / Microsoft | Phi-3-mini-4k-instruct-generic-cpu:3; -cuda-gpu:2 (version skew) | UNKNOWN; 3.8B params | UNKNOWN | available |
phi-3-mini-128k | Microsoft / Microsoft | Phi-3-mini-128k-instruct-generic-cpu:3; -cuda-gpu:2 | UNKNOWN; 3.8B params | UNKNOWN | available |
gpt-oss-20b | Microsoft / OpenAI (open weight) | gpt-oss-20b-generic-cpu:1; -cuda-gpu:1 | UNKNOWN on ONNX | UNKNOWN | available |
qwen3-0.6b | Microsoft / Alibaba | qwen3-0.6b-generic-cpu:4; -cuda-gpu:2 | UNKNOWN | UNKNOWN | available |
qwen3-1.7b | Microsoft / Alibaba | qwen3-1.7b-generic-cpu:2; -cuda-gpu:2 | UNKNOWN | UNKNOWN | available |
qwen3-4b | Microsoft / Alibaba | qwen3-4b-generic-cpu:3; -cuda-gpu:2 (version skew) | UNKNOWN | UNKNOWN | available |
qwen3-8b | Microsoft / Alibaba | qwen3-8b-generic-cpu:2; -cuda-gpu:2 | UNKNOWN | UNKNOWN | available |
qwen3-14b | Microsoft / Alibaba | qwen3-14b-generic-cpu:2; -cuda-gpu:2 | UNKNOWN | UNKNOWN | available |
qwen3.5-2b-text | Microsoft / Alibaba | qwen3.5-2b-text-generic-cpu:1; -cuda-gpu:1 | UNKNOWN | UNKNOWN | available |
qwen2.5-0.5b | Microsoft / Alibaba | qwen2.5-0.5b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN; 0.5B params | UNKNOWN | available |
qwen2.5-1.5b | Microsoft / Alibaba | qwen2.5-1.5b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN; 1.5B params | UNKNOWN | available |
qwen2.5-7b | Microsoft / Alibaba | qwen2.5-7b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
qwen2.5-14b | Microsoft / Alibaba | qwen2.5-14b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
qwen2.5-coder-0.5b | Microsoft / Alibaba | qwen2.5-coder-0.5b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
qwen2.5-coder-1.5b | Microsoft / Alibaba | qwen2.5-coder-1.5b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
qwen2.5-coder-7b | Microsoft / Alibaba | qwen2.5-coder-7b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
qwen2.5-coder-14b | Microsoft / Alibaba | qwen2.5-coder-14b-instruct-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
deepseek-r1-7b | Microsoft / DeepSeek | deepseek-r1-distill-qwen-7b-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
deepseek-r1-14b | Microsoft / DeepSeek | deepseek-r1-distill-qwen-14b-generic-cpu:4; -cuda-gpu:4 | UNKNOWN | UNKNOWN | available |
mistral-7b-v0.2 | Microsoft / Mistral AI | mistralai-Mistral-7B-Instruct-v0-2-generic-cpu:3; -cuda-gpu:2 | UNKNOWN | UNKNOWN | available |
mistral-nemo-12b-instruct | Microsoft / Mistral AI and NVIDIA | mistral-nemo-12b-instruct-generic-cpu:1; -cuda-gpu:1 | UNKNOWN | UNKNOWN | available |
olmo-3-7b-instruct | Microsoft / Allen Institute for AI | olmo-3-7b-instruct-generic-cpu:1; -cuda-gpu:1 | UNKNOWN | UNKNOWN | available |
smollm3-3b | Microsoft / Hugging Face | smollm3-3b-generic-cpu:1; -cuda-gpu:1 | UNKNOWN | UNKNOWN | available |
whisper-tiny | Microsoft / OpenAI | openai-whisper-tiny-generic-cpu:4; -cuda-gpu:3 | UNKNOWN | UNKNOWN | available (ASR) |
whisper-base | Microsoft / OpenAI | openai-whisper-base-generic-cpu:3; -cuda-gpu:3 | UNKNOWN | UNKNOWN | available (ASR) |
whisper-small | Microsoft / OpenAI | openai-whisper-small-generic-cpu:4; -cuda-gpu:3 | UNKNOWN | UNKNOWN | available (ASR) |
whisper-medium | Microsoft / OpenAI | openai-whisper-medium-generic-cpu:4; -cuda-gpu:3 | UNKNOWN | UNKNOWN | available (ASR) |
whisper-large-v3-turbo | Microsoft / OpenAI | openai-whisper-large-v3-turbo-generic-cpu:4; -cuda-gpu:3 | UNKNOWN | UNKNOWN | available (ASR) |
nemotron-speech-streaming-en-0.6b | Microsoft / NVIDIA | nemotron-speech-streaming-en-0.6b-generic-cpu:3; -cuda-gpu:1 | UNKNOWN; 0.6B params | UNKNOWN | available (ASR) |
nemotron-speech-streaming-es-0.6b | Microsoft / NVIDIA | nemotron-speech-streaming-es-0.6b-ft-generic-cpu:1; -cuda-gpu:1 | UNKNOWN; 0.6B params | UNKNOWN | available (ASR) |
nemotron-3.5-asr-streaming-0.6b | Microsoft / NVIDIA | nemotron-3.5-asr-streaming-0.6b-generic-cpu:3; -cuda-gpu:2 | UNKNOWN; 0.6B params | UNKNOWN | available (ASR) |
What "available" means here
The entry comes from the foundry-local source that the device SDK also consumes, and the device documentation names several of these aliases directly. It does not mean each row has been observed in a device-side foundry model list. Confirming that needs a running install, and it is recorded as an open unknown in SPIKE-22.
Two further device-side entries appear in the CLI reference as sample filter values but not in the catalog snapshot: deepseek-r1-distill-qwen-1.5b-generic-cpu and the alias phi4-cpu. Those are documentation samples rather than a catalog dump, so they suggest more breadth without proving it.
Section 2: NPU and alternate-accelerator variants, exclusive to this target
Azure Local Foundry's catalog sync paginates CPU and GPU only, so every variant class below is absent from it. These are what you gain by running on a device rather than a cluster.
| Variant class | Suffix | Execution provider | Hardware |
|---|---|---|---|
| Qualcomm NPU | -qnn-npu | QNNExecutionProvider | Snapdragon X Elite / X Plus, driver 30.0.140.0+ |
| AMD NPU | -vitis-npu | VitisAIExecutionProvider | Adrenalin 25.6.3 to 25.9.1 |
| Intel NPU and GPU | (per model) | OpenVINOExecutionProvider | Intel TigerLake+ / AlderLake+ / ArrowLake+ |
| Generic GPU (WebGPU/Dawn) | -generic-gpu | WebGpuExecutionProvider | the cross-platform GPU fallback |
| NVIDIA TensorRT RTX | (per model) | NvTensorRTRTXExecutionProvider | RTX 30 series and newer |
The CPU execution provider is always available as a fallback. If no GPU or NPU is detected, inference runs on CPU automatically.
The specific per-model NPU list is not enumerable from documentation. Microsoft's own guidance is to ask the hardware:
bash
foundry model list --filter device=NPU
foundry model list --filter task=chat-completion
foundry model info phi-4-miniThat command is the authority, and this page is not. A capability read from documentation is a claim; exercised against the runtime it is a fact.
Compile your own
You are not limited to the catalog. Models compiled to ONNX can be run on this target, including Hugging Face models converted for the purpose. See Compile Hugging Face models to run on Foundry Local.
Keeping this current
This page is a transcription of a catalog snapshot dated 2026-07-28. It will drift, and the product is in active development. Generating it from the live catalog is tracked as a feature request rather than left as a recurring manual chore: see issue #15, auto-refresh the model availability catalogs.
See also
- Model availability matrix, all three targets side by side, filterable.
- Available models: Azure AI Foundry, the hosted cloud roster.
- Available models: Azure Local Foundry, the cluster roster.
- SPIKE-22, the research behind every row here, including the open unknowns.
- Model catalog, what this project chose and why.
- Model selection, the methodology behind the choosing.
- Foundry Local: models.