Microsoft releases a 53GB local build of MAI-Code-1.1-Flash
A 3-bit quantized build keeps a 256K context and approaches the cloud model’s SWE-bench score, but needs more than 120GB of RAM and lacks clearly published weight-license terms.
2 min read
Microsoft released an on-device build of MAI-Code-1.1-Flash on October 7, extending its coding model beyond hosted GitHub Copilot inference. The local version is a quantized mixture-of-experts model with 137 billion total parameters and 6.8 billion active parameters. Microsoft says it preserves the full 256,000-token context window and can be selected directly or used by Copilot’s forthcoming local/cloud router.
The practical trade-off is hardware. Microsoft recommends more than 120GB of RAM; its Surface Laptop Ultra test used a 53GB model build and peaked at 75.5GB of memory at 256K context. The company reports prompt-processing throughput of 923.5 tokens per second at 64K context and 769.8 tokens per second at 128K, using an ARM64 llama.cpp CUDA runtime. Local calls carry no per-inference charge, but the Copilot integration remains experimental and is due across the Copilot app, CLI and VS Code by the end of October.
Microsoft’s release evaluation compares its roughly 3.3-bit local build with the cloud BF16 version and an Unsloth GPT-OSS-120B GGUF baseline. The local model scored 70.8% on SWE-bench Verified’s 500 tasks, versus 72.6% for the cloud build, and 66.29% on the 89-task Terminal-Bench 2.1 set, versus 62.9% for the cloud build. Those are vendor-run results from an October 5 configuration, not independent tests; Terminal-Bench’s small sample also makes broad conclusions risky.
The announcement says the model is downloadable, but Microsoft’s public launch materials do not clearly state the weight license or provide a durable public weights link. Until those terms are explicit, it should be treated as a locally runnable proprietary model—not an open-source release.
Sources: Microsoft AI announcement · Microsoft technical deep dive · Model page · Unsplash image license
Conversation
0 approved commentsNo comments yet. Start a thoughtful conversation.
Join the conversation.
Sign in with ChatGPT to leave a comment or rate this article.