{"version":1,"type":"story","url":"https://digestai.news/story/microsoft-foundry-local-runs-qwen3-8-27b-locally-with-github-copilot-o","json":"https://digestai.news/story/microsoft-foundry-local-runs-qwen3-8-27b-locally-with-github-copilot-o.json","markdown":"https://digestai.news/story/microsoft-foundry-local-runs-qwen3-8-27b-locally-with-github-copilot-o.md","slug":"microsoft-foundry-local-runs-qwen3-8-27b-locally-with-github-copilot-o","headline":"Microsoft Foundry Local runs Qwen3.8-27B locally with GitHub Copilot on Nvidia GPUs","summary":"Microsoft released Foundry Local CLI 0.13.0, which runs the Qwen3.8-27B model on Nvidia GPUs and connects it to GitHub Copilot via Bring Your Own Key. The 4-bit quantized model needs about 18 GB VRAM, so it fits on consumer cards with 24 GB or more. On a GeForce RTX 5090 with 32 GB VRAM, Foundry Local showed higher token generation speed than llama.cpp at every prompt length tested, with the gap widening past 131K tokens. GPUs with 32 GB VRAM support up to 262K tokens of context; 24 GB cards support up to 64K. The stack uses DFlash2 speculative decoding, CUDA graph fusions, KV-cache quantization, prefix caching and continuous batching. Microsoft says local models may have limited reasoning versus advanced cloud models, so a hybrid workflow is suggested. Mac and AMD GPU support are planned next.","keyPoints":["Foundry Local CLI 0.13.0 runs Qwen3.8-27B locally and plugs into GitHub Copilot BYOK","4-bit model needs ~18 GB VRAM; 24 GB cards support 64K context, 32 GB cards support 262K","Microsoft benchmarks show faster generation than llama.cpp on RTX 5090, especially past 131K tokens"],"whyItMatters":"Developers can run a capable coding agent on their own GPU, keeping code private and avoiding cloud costs or latency.","category":{"slug":"agents","name":"Agents & Tools","url":"https://digestai.news/category/agents"},"entities":{"companies":["Microsoft","GitHub","Nvidia"],"models":["Qwen3.8-27B"],"people":[]},"firstPublishedAt":"2026-11-06T05:14:39Z","updatedAt":"2026-11-06T05:14:39Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"devblogs.microsoft.com","title":"Bring Agentic Coding On-Device with Foundry Local","url":"https://devblogs.microsoft.com/foundry/foundry-local-agentic-coding-qwen3-8-27b","publishedAt":"2026-11-06T05:14:39Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1xp4lv5/bring_agentic_coding_ondevice_with_foundry_local/","points":null}],"thread":null,"cite":{"text":"Digest AI, \"Microsoft Foundry Local runs Qwen3.8-27B locally with GitHub Copilot on Nvidia GPUs\", 6 November 2026, https://digestai.news/story/microsoft-foundry-local-runs-qwen3-8-27b-locally-with-github-copilot-o","publisher":"Digest AI","title":"Microsoft Foundry Local runs Qwen3.8-27B locally with GitHub Copilot on Nvidia GPUs","datePublished":"2026-11-06T05:14:39Z","url":"https://digestai.news/story/microsoft-foundry-local-runs-qwen3-8-27b-locally-with-github-copilot-o"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}