Microsoft AI releases MAI-Code-1.1-Flash at a quarter of 1.0's price, live in GitHub Copilot, with a 3-bit on-device version

Why it matters

【Key points】

Microsoft AI released MAI-Code-1.1-Flash on Oct 7, 2026 (02:05 Oct 8 Beijing time). Versus 1.0 (Build, June) it claims better code, 25% better token efficiency and a quarter of the price; it is in production in GitHub Copilot. GitHub lists it as GA at $0.20/M input, $0.02 cached, $1.20 output.

A 3-bit on-device version keeps the 256K context and is "comparable" to full precision (no figures). Local calls, picked by Copilot's router, carry no inference charge; 120GB+ RAM is recommended. Microsoft says it is "available to download and run locally" but names no channel or license. Experimental Copilot app/CLI/VS Code access comes by end of October.

【Verifiable】

https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/ ; https://x.com/MicrosoftAI/status/2107926901150908483 . Self-reported gains over 1.0: Terminal-Bench 2.1 in Copilot CLI +22%, .NET tasks +15%, code survival +4%, return visits +9%, streaming 25% faster, 25% fewer tokens. Prices: https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing (GitHub notes MAI checkpoints keep changing). The microsoft/MAI-Code repo has docs only (MIT), no weights, and nothing is on Hugging Face, so we don't call it open. Claude Haiku 5.5 joined Copilot the same day at $0.10/$0.50 up to 100K tokens. NVIDIA's blog gives RTX Spark up to 128GB unified memory.

【Our read】

1. "A quarter" is versus Microsoft's own 1.0. On the same Copilot list Haiku 5.5 is cheaper up to 100K tokens. Microsoft's edge is owning the router and a local option.

2. Free local inference shifts compute cost to users' hardware. The 120GB bar happens to fall within RTX Spark's 128GB maximum; a fit, not proof of design intent; few machines qualify.

3. All metrics are relative to undisclosed bases and checkpoints change; hard to reproduce.

4. Test: will a download channel and license appear by end of October? Until then the local version is a Copilot component, not an open model.