MAI-Code-1.1-Flash
Microsoft's 137B coding MoE you can run locally — 3-bit quantized, 256K context, no cloud token spend.
About
MAI-Code-1.1-Flash is Microsoft AI's coding-focused mixture-of-experts model: 137 billion total parameters with 6.8 billion active per token. The cloud model shipped inside GitHub Copilot on August 11, 2026; the news is the local build Microsoft announced on October 7, 2026 at its Windows and Surface event — downloadable, runnable on your own hardware, with no per-call inference charge for local use.
The local version uses ~3-bit mixed-precision quantization plus speculative decoding: a 53 GB model footprint (about 80% smaller than the Bfloat16 cloud variant), 75.5 GB peak memory at the full 256,000-token context, and Microsoft recommends devices with more than 120 GB of RAM for best performance — the reference numbers come from a Surface Laptop Ultra with up to 128 GB of unified memory. Microsoft reports the quantized build at 70.8% on SWE-Bench Verified (cloud: 72.6%) and 66.29% on Terminal-Bench 2.1 (cloud: 62.9%), i.e. comparable coding performance at a fraction of the footprint.
Experimental on-device access in GitHub Copilot (app, CLI, VS Code) is planned for end of October 2026, with Copilot routing between local and cloud. Code execution runs inside Microsoft's Executable Container (MXC) sandbox.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported (Microsoft): 3-bit local build 70.8% SWE-Bench Verified (cloud 72.6%), 66.29% Terminal-Bench 2.1 (cloud 62.9%). Independent: no third-party rerun.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Release facts (October 7, 2026 local announcement; cloud in Copilot since August 11; 137B/6.8B MoE; 53 GB footprint; 75.5 GB peak at 256K; >120 GB RAM recommendation) corroborated across Microsoft's announcement and multiple independent outlets. Docked for: (1) all benchmark figures are Microsoft's own with no independent rerun; (2) the hardware bar excludes nearly all existing laptops; (3) downloadable but proprietary — not open weights, and local pricing terms beyond 'no inference fee' are unpublished.