01 What happened
On November 5, 2026, Microsoft announced Foundry Local, a platform for on-device agentic coding running the Qwen3.8-27B model locally. Code and context remain on the device, reducing cloud costs, improving privacy, and enabling offline work, although cloud models are recommended for complex reasoning tasks. It requires an NVIDIA GPU with at least 24 GB VRAM.
02 Key details
- Foundry Local runs Qwen3.8-27B locally on NVIDIA GPUs with at least 24 GB VRAM; 24 GB supports up to 64K token context, and 32 GB or more reaches the model's maximum 262K tokens.
- Code and context never leave the device; offline work is supported. Integrates with GitHub Copilot via Bring Your Own Key (BYOK).
- Microsoft benchmarked Foundry Local against llama.cpp on an RTX 5090 (32 GB) and DGX Spark; Foundry Local achieved higher generation speed at every tested prompt length, with the advantage widening beyond 131K tokens.
- Requires Foundry Local CLI 0.13.0+; CUDA and cuDNN must be installed separately as Foundry Local does not download these dependencies. Expansion to Macs and AMD GPUs is planned.
03 Why it matters
This gives developers access to agentic AI coding locally, avoiding cloud API costs, keeping sensitive code on-device, and working without internet, while accepting limited reasoning compared to cloud models.
04 Who it matters to
Software developers.
Original source