Generative small language models in Foundry Local

Foundry Local on Azure Local includes a curated catalog of generative small language models (SLMs) that you can deploy for on-premises inference. These models are optimized to run on constrained hardware while delivering strong performance on tasks like chat completion, code generation, and reasoning. Depending on the model entry you choose, generative inference uses either ONNX Runtime or vLLM.

This article provides guidance on choosing the right model for your workload.

Important

  • Foundry Local is available in preview. Preview releases provide early access to features that are in active deployment.
  • Features, approaches, and processes can change or have limited capabilities before general availability (GA).

What are small language models?

Small language models are compact generative AI models, typically ranging from under 1 billion to around 14 billion parameters. Compared to large language models (LLMs) with hundreds of billions of parameters, SLMs offer:

  • Lower resource requirements — Run on a single GPU or even on CPU, making them practical for edge and on-premises environments.
  • Faster inference — Smaller model sizes translate to lower latency and higher throughput per device.
  • Local deployment — Keep data processing on-premises where your data is generated, without relying on cloud endpoints.

SLMs trade some breadth of general knowledge for efficiency, but advances in training techniques mean that modern SLMs can match or exceed larger models on focused tasks.

Choose a model

The Foundry Local catalog includes generative models from different providers. All models support the chat completion task and use OpenAI-compatible REST API patterns. The CC value in the table refers to the NVIDIA GPU compute capability version.

Use Inference runtimes in Foundry Local on Azure Local to compare ONNX Runtime and vLLM for hardware requirements and performance characteristics.

The following table highlights representative examples only. For the complete and most current list, see Foundry Local model catalog.

Model Publisher Max context length Recommended minimal GPU generation Required GPU memory
Phi-3.5-mini-instruct Microsoft 29,472 Ampere (CC 8.0)+ 8.428 GB
Phi-4-mini-instruct Microsoft 93,520 Ampere (CC 8.0)+ 7.806 GB
Phi-4-mini-reasoning Microsoft 93,520 Ampere (CC 8.0)+ 7.806 GB
Mistral-7B-Instruct-v0.2 Mistral AI 29,328 Ampere (CC 8.0)+ 15.64 GB
gpt-oss-20b OpenAI 96,784 Blackwell (CC 10.0)+ 14.793 GB

For GPU-specific settings and performance benchmarks, see vLLM model reference.

Next steps