Qwen3.8-27B now runs fully local in GitHub Copilot via Foundry Local
An Apache-2.0 27B coder with 262K context, no cloud and no per-token bill. Plus Mistral's managed-agents preview and a Nov 30 Claude deprecation clock.

Copy markdown
Qwen3.8-27B goes fully offline inside Copilot
Microsoft's Foundry Local (CLI 0.13.0+) serves Alibaba's Apache-2.0 Qwen3.8-27B on a single 24GB-VRAM GPU — about 18GB at 4-bit — and adds it to GitHub Copilot as a BYOK model provider, so your code and context never leave the machine. Microsoft reports 61.7 on SWE-bench Pro and 73.0 on Terminal Bench 2.1, with a 64K context window on 24GB and the full 262K on 32GB and up.
Mistral opens managed agents in Studio
Mistral Studio's managed agents are in public preview: multi-step agents with reusable skills, code sandboxes, subagents and hooks, plus resumable workflows you can generate from a prompt in Vibe and ship as apps with one command via the new Mistral MCP server. They run on the ~1T-param Large 4 'Le Chonk,' and the whole platform can sit in your own VPC or on-prem.
Deprecation clock: two Claude models change Nov 30
Mark Nov 30: Anthropic removes fast mode for claude-opus-4-8 (speed:"fast" requests will start erroring — migrate to Opus 5.5) and retires claude-sonnet-4-5-20250929 from the API, pointing users to Sonnet 5.5. Both are one-line model-string swaps, but calls break if you wait.