What Microsoft Announced
MAI-Code-1.1-Flash is the coding-focused mixture-of-experts model already running in GitHub Copilot in the cloud (since August 11, 2026). What is new is a local build, announced October 7, 2026, that developers can download and run on their own Windows hardware — no active internet connection required, demonstrated on stage with Airplane mode on. Local model calls carry no inference charges.
The Specs and the Catch
The model has 137 billion total parameters with 6.8 billion active, and keeps its full 256,000-token context window after 3-bit quantization. Microsoft's numbers: a 53 GB model footprint, 75.5 GB peak memory use at full context — and a recommendation of more than 120GB of RAM for best performance, because the OS, apps, and inference runtime share the memory pool. This is not for an ordinary laptop.
- Model: MAI-Code-1.1-Flash, 137B total / 6.8B active MoE
- Context: 256K tokens, full window kept locally
- Footprint: 53 GB quantized, 75.5 GB peak memory
- RAM: Microsoft recommends more than 120GB
- Cost: zero per-call inference fees for local use
Performance: Local vs Cloud
The shrinking barely costs anything. In Microsoft's testing on SWE-bench Verified (built from real GitHub issues), the quantized on-device version resolved 70.8% of tasks versus 72.6% for the higher-precision cloud variant — and it even scored higher than the cloud model on one of Microsoft's two coding benchmarks. On a Surface Laptop Ultra with Nvidia RTX Spark, it generated roughly 60 tokens per second on short prompts, dropping to just under 40 at a 256K-token prompt.
The Copilot Roadmap
Experimental on-device access inside GitHub Copilot — across the GitHub Copilot app, Copilot CLI, and Visual Studio Code — is planned for the end of October 2026, so Copilot cannot use the local model yet. The pitch is a hybrid: the PC handles routine coding-agent loops locally while the cloud stays available for heavier work, moving agent work off per-token cloud billing and onto desk-side hardware.
Frequently Asked Questions
What is MAI-Code-1.1-Flash?
It is Microsoft's coding-focused mixture-of-experts model (137B total, 6.8B active parameters, 256K context), already used in GitHub Copilot in the cloud. On October 7, 2026, Microsoft announced a downloadable 3-bit-quantized version that runs entirely on a Windows PC.
Is running MAI-Code-1.1-Flash locally free?
Local calls carry no per-call inference fees, but that covers inference charges only — not the cost of the hardware or electricity. You still need a machine with very large memory.
How much RAM does MAI-Code-1.1-Flash need?
Microsoft recommends more than 120GB of RAM for best performance. The quantized model has a 53 GB footprint and up to 75.5 GB peak memory use at its full 256K-token context. Only top-spec machines like the 128GB Surface Laptop Ultra clear the bar.
When will GitHub Copilot use the local model?
Experimental on-device access in GitHub Copilot (app, CLI, and VS Code) is planned for the end of October 2026. Until then, the download is for direct local use, not wired into Copilot.
