The Pulse
Microsoft Takes Its 137B MAI Coding Model Local in Copilot
Microsoft says GitHub Copilot will run a quantized version of its 137-billion-parameter MAI Code 1.1 Flash model on supported local hardware. Experimental Copilot access is planned for the app, CLI and Visual Studio Code by the end of Octob
AI.info Team · Written with AI ·
Microsoft says it has adapted a 137-billion-parameter coding model to run on local computers, with GitHub Copilot set to route some coding work between on-device and cloud models. The announcement, published October 7, names MAI Code 1.1 Flash as a mixture-of-experts model with 6.8 billion active parameters; the Copilot integration is planned as an experimental release by the end of October.
The hardware demands narrow the immediate audience. Microsoft recommends more than 120GB of RAM for local use and reports its first shipping configuration on a Surface Laptop Ultra equipped with NVIDIA RTX Spark and up to 128GB of unified memory. The company also says the local model can be downloaded and run independently, with no per-inference charge for local calls.
Copilot will choose between local and cloud
Microsoft plans to offer two ways to use local models across the GitHub Copilot app, CLI and Visual Studio Code: automatic routing between local and cloud inference, or explicit selection of a local model. In automatic mode, Copilot will weigh task context and cached work when deciding where to send a task. Microsoft says the feature is due by the end of October, not generally available on October 7.
“To bring this local-development experience to life, Microsoft AI developed a local version of MAI Code 1.1 Flash, a coding-optimized mixture-of-experts model with 137 billion total and 6.8 billion active parameters,” write Patrick Nikoletich, distinguished product manager at GitHub, and Stuart Schaefer, partner architect for Windows Platform at Microsoft, in Microsoft’s announcement.
Quantization cuts the model to 53GB
Microsoft uses quantization and speculative decoding to reduce the model’s memory footprint and improve response speed. It says the on-device version is about 53GB—an 80% reduction from its bfloat16 cloud counterpart—and uses mixed-precision quantization at approximately 3.3 bits per weight. At a 256,000-token context, the company reports peak memory use of 75.5GB, leaving the operating system, applications and other processes to share the machine’s remaining capacity.
Microsoft’s published tests put the quantized version at 70.80% on SWE-Bench Verified, compared with 72.6% for its full-precision MAI Code 1.1 Flash, and 66.29% on Terminal-Bench 2.1, compared with 62.9% for the full-precision model. The tests were run on October 5 with a Windows ARM64 llama.cpp CUDA runtime and DFlash2 speculative decoding. Microsoft warns that actual results vary by device and configuration; the figures are company-reported benchmark results, not an independent assessment.
Local inference does not mean offline
Copilot’s planned routing feature is designed to keep cloud models available for tasks that need them, so using local inference does not make a session offline. Microsoft separates where a model runs from what its coding agent can access: sandbox policies govern the agent’s tools and commands, not the model’s location.
The announcement describes Microsoft Execution Containers as the mechanism for restricting access to files, networks, credentials and system capabilities. It also sets limits on that protection: built-in file tools are checked by Copilot’s agent harness rather than isolated as operating-system child processes, and remote Model Context Protocol servers sit outside the local process sandbox. The practical offer is therefore a large local coding model paired with optional, separately configured boundaries around agent actions—not a fully disconnected or uniformly isolated Copilot session.