Microsoft has introduced a local edition of MAI-Code-1.1-Flash that runs on Windows PCs without server access. The model requires substantial memory and is coming to GitHub Copilot in experimental preview by the end of October, starting with Nvidia RTX Spark PCs.

Wrong: Claude Opus 5.5's 87.6% and GPT-6 Astra's 87.3% Terminal-Bench 2.1 scores read as Microsoft's own figures. Right: those scores come from Vals AI's Terminal-Bench 2.1 leaderboard, which uses a different test setup; Microsoft reported only its own model's scores (62.9% cloud, 66.3% local).
Wrong: said Microsoft recommends more than 120 GB of video memory to run the local MAI-Code-1.1-Flash. That figure came from Frandroid, not Microsoft. Right: Microsoft says the local model's memory use peaks at 75.5 GB at its full 256K context, so it effectively needs a 128 GB RTX Spark machine.
Wrong: said MAI-Code-1.1-Flash activates 5 billion parameters per step and was "confirmed in development" for GitHub Copilot, and gave only the cloud model's 62.9% Terminal-Bench 2.1 score. Right: Microsoft says the model has 137 billion total and 6.8 billion active parameters; the local version comes to GitHub Copilot in experimental preview by the end of October; on Terminal-Bench 2.1 the cloud model scored 62.9% and the local version 66.3%.
Microsoft says memory use peaks at 75.5 GB at the full 256K context, so it effectively needs a 128 GB RTX Spark machine.
Computations occur on the device itself, so Microsoft states users will not be billed for tokens.
Tap a lens to see what this story means for you.
Liked this? The Brief brings you the whole day in tech, verified, every morning.
Two minutes, free forever. What's in The Brief →
See what’s happening right now
The Feed runs all day — short, verified briefs the moment they break.
Open the FeedFollow @thecircuitry_
Every story we publish, as it happens. No noise between.
Reader-supported
The Circuitry is a passion project I've always wanted to build, and I love the work behind it.
Running it costs real money. APIs, hosting, time. To keep improving the site and growing this into something useful for everyone, those costs have to be covered.
Any contribution is appreciated. If not, no pressure. Thanks for reading.
OpenAI has cut the average success rate of attacks that extract protected reasoning from an earlier request to 0.2% on GPT-6.2 Astra, from 54.0% on GPT-6 Astra. Current-request attacks are at 0.9%, and selected safeguard improvements are being applied back to GPT-6 Astra.
OpenAI has introduced MedChemBench, a 99-problem test of how models read molecules, explain lab results, and choose the next compound. At the xhigh setting, GPT-6.2 Astra scored 84.5%, while compound choice and property forecasts remain much weaker than structure work.
Autonomous Cairn agents found an unpatched flaw in a medical imaging app and took control of a server in a university campaign that began September 8, 2026. Amazon rates a Chinese-nexus link as moderate confidence, and the project advisory still lists no patched version.
TikTok has launched a public AI checker for images and videos made with its own tools and has started a global opt-in likeness rollout. The checker is live without an account, while the likeness tool is limited to eligible creators and a notice does not by itself remove a post.
NVIDIA has reported third-quarter revenue of $115.8 billion, up 103% from a year ago, with data-center revenue of $107.6 billion. Fourth-quarter revenue is expected at $130.0 billion, plus or minus 3%, with no data-center compute revenue from China assumed.