How Microsoft Put A 137B Coding AI On A Laptop (Run AI Locally)
Microsoft put a 137B coding model on a Surface Laptop Ultra: what its 70.8% SWE-bench score, its 75.5 GB memory peak and its unverified download mean for local AI coding Microsoft adapted MAI-Code-1.1-Flash, a 137-billion-parameter sparse mixture-of-experts model with 6.8 billion parameters active per token, to run locally on a Surface Laptop Ultra. Microsoft reports the 53 GB quantized build (mixed precision, about 3.3 bits per weight) resolved 70.8% of SWE-bench Verified's 500 issue-repair tasks, against 72.6% for the cloud model, and 66.29% versus 62.9% on Terminal-Bench 2.1's 89 terminal tasks. These are vendor-reported numbers: they show the compressed model kept most of its tested capability under Microsoft's harness, not that compression made it smarter or that it will behave the same on your own codebase. The file size is not the memory budget. Microsoft measured a peak of 75.5 GB at a 256K-token context for the model session, and your operating system, editor, browser and build tools come on top; Microsoft recommends more than 120 GB of RAM for best performance. The 923.5 tokens per second at 64K is prompt-reading (prefill) speed, not the time to a passing fix: speculative decoding speeds up generation, but tool runs, test failures and retries are most of an agent session. What you can use today: the cloud model has been in GitHub Copilot since August 11. Microsoft says the local edition is available to download, but our own checks on October 8 did not verify a runnable download or the license for the weights. Copilot CLI 1.0.94 can discover models you already run in Ollama (they must support tool calling and streaming), and Copilot's integrated local and cloud routing is experimental, due by the end of October. On cost, local calls carry no per-token fee (the cloud rate is $0.20 per million input tokens and $1.20 per million output tokens), but the 128 GB Surface Laptop Ultra configuration lists at $5,899.99, so buying hardware just to cut token bills is an unproven bet. And a local model does not make your session offline: choosing one doesn't turn on offline mode or disable GitHub telemetry, and your tools can still reach the network. The answer: if you already own a machine with more than 120 GB of memory, test it on one real repair once working weights and runtimes exist; everyone else should keep their current cloud or smaller local workflow. Image credits: "If it fits, I sits" cat photo by Big Ben in Japan (CC BY-SA 2.0); banana photo by Evan-Amos (CC BY-SA 3.0). Chapters: 0:00 Intro 1:36 Code 3:43 Memory 5:58 Speed 8:45 Access 11:04 Cost 14:21 Control 16:29 Conclusion Tools & resources mentioned: - GitHub Copilot: https://github.com/features/copilot - GitHub Copilot CLI: https://github.blog/changelog/2026-10-07-discover-local-models-in-github-copilot-cli/ - Surface Laptop Ultra: https://www.microsoft.com/en-us/surface/devices/surface-laptop-ultra - MAI-Code repository: https://github.com/microsoft/MAI-Code - Ollama: https://ollama.com - llama.cpp: https://github.com/ggml-org/llama.cpp About The Stack The Stack helps you build with AI. Each video takes one tool, model, or workflow and shows how it works in a few focused minutes, with the real benchmarks and real costs. We go deep on Claude Code and Cursor for AI coding, AI agents and MCP servers, the open-source AI tools and GitHub repos most people miss, RAG and vector search, fine-tuning, and running local LLMs on your own machine with Ollama and LM Studio. We compare models like ChatGPT and Claude, test AI automation with Zapier, Make, and n8n, and flag the tools that actually ship. Subscribe for new breakdowns: https://www.youtube.com/@the-stack-ai?sub_confirmation=1 #local AI coding #GitHub Copilot #coding agents
