MAI-Code-1.1-Flash downloadable for on-device coding

Microsoft released its 137B-total / 6.8B-active coding model as a downloadable ~3-bit quantized build with full 256K context and zero inference charges for local calls.

Why it matters

What you could build with it

Does it hold up?

Too early to judge the downloadable on-device build: it was announced Oct 7 and Copilot experimental access only rolls out 'by end of the month'. The 3-bit quant keeps full 256K context and matches the full model on SWE-Bench Verified/Terminal-Bench 2.1 per Microsoft — but all evidence is vendor-reported and >120GB RAM is the stated floor. No independent hands-on runs exist yet.

Built with MAI-Code-1.1-Flash downloadable for on-device…

• VS Code demo: shipping a real feature end-to-end with MAI-Code-1-Flashyoutube · Walkthrough of MAI-Code-1-Flash inside VS Code Copilot Chat: codebase exploration, building a feature, running and testing it, plus a cost-benefit breakdown. (Note: published for the 1.0 launch, but shows the on-device/Copilot router concept now extended to 1.1.)
• Neowin: Microsoft releases MAI-Code-1.1-Flash to compete with Chinese modelsneowin.net · 1.1 pricing detail: 0.25x premium-request multiplier for annual Copilot subscribers ($0.20/$0.02/$1.20 per 1M input/cached/output), 22% better on Terminal-Bench 2.1 in Copilot CLI, 15% better .NET tasks, native vision support — framed as Microsoft's answer to GLM-5.2 and Kimi K3.

Learn more

announcement →

First spotted on article: source.

More AI models