Microsoft 将 Copilot 例行编码任务切换至本地模型 MAI-Code-1.1-Flash
Copilot 的杂活以后不用云端跑 Claude Haiku 4.5 了，改用微软自家 6.8B 激活参数的小模型在你电脑上跑，本地调用不收费，还保住了 256K 上下文。
Microsoft 宣布 Copilot 的例行编码工作改由自研模型 MAI-Code-1.1-Flash 在本地执行，取代 Anthropic 的 Claude Haiku 4.5。该模型总参数 137B、激活参数 6.8B，采用 3-bit 量化，仍保留 256K token 上下文窗口，官方称在 SWE-Bench Verified 和 Terminal-Bench 2.1 上与全精度成绩相同。模型路由器会把简单任务留在机器上运行，本地调用不产生推理费用。硬件方面依赖搭载 Nvidia RTX Spark 芯片的 Surface Laptop Ultra（2599 美元起）和 Surface RTX Spark Dev Box（5999 美元起）。同时 Microsoft 的 agent 沙箱 Microsoft Execution Containers（MXC）在 Windows 11 上正式可用，支持从独立进程到虚拟机的多级隔离。
Microsoft just announced that Copilot will hand routine coding from Anthropic's Claude Haiku 4.5 to Microsoft's MAI-Code-1.1-Flash (a 137B total and 6.8B active param model) running on Nvidia-powered PCs
A model router will keep th routine work on the machine and this local model calls carry no inference charge.
The on-device MAI-Code-1.1-Flash uses 3-bit quantization yet keeps its 256K-token context window, and Microsoft says it scores the same as full precision on SWE-Bench Verified and Terminal-Bench 2.1.
Microsoft's hardware for the shift is the Surface Laptop Ultra from $2,599 and the Surface RTX Spark Dev Box from $5,999
Both run on Nvidia's RTX Spark, an Arm-based chip with a Blackwell GPU and up to 128GB of memory shared by processor and graphics.
Also, Microsoft made its agent sandbox, Microsoft Execution Containers (MXC), generally available on Windows 11. Companies can now set which files and network locations an agent may use, and Windows blocks everything else while the agent runs, with isolation levels ranging from separate processes and sessions up to virtual machines and cloud PCs.
