Microsoft takes MAI-Code-1.1-Flash local: 137B MoE scores 70.8% SWE-Bench Verified on-device, GitHub Copilot to route between local and cloud

At its Oct 7 Windows event Microsoft introduced 'hybrid intelligence': run agent work locally when it makes sense and reach the cloud for frontier capability, with no inference charge for local calls. MAI-Code-1.1-Flash, a 137B-total / 6.8B-active MoE coding model, now ships in an on-device build using ~3.3 bits-per-weight mixed-precision quantization (nearly 80% smaller) and DFlash2 sliding-window speculative decoding on a Windows ARM64 llama.cpp CUDA runtime, keeping a 256K context. In Microsoft's Oct 5 tests on Surface Laptop Ultra it scored 70.80% on SWE-Bench Verified (cloud 72.6%) and 66.29% on Terminal-Bench 2.1 (cloud 62.9%), versus 32.0% / 23.6% for GPT-OSS-120B, with 75.5GB peak memory at 256K and 923.5 / 769.8 tok/s prompt processing at 64K / 128K; these are vendor numbers without independent replication yet. GitHub's HydraFusion orchestrator will route Copilot tasks between local and cloud models in experimental preview later in October across the Copilot app, Copilot CLI and VS Code. Available today: Copilot CLI 1.0.94-0 discovers local Ollama models via /model (tool calling and streaming required; offline mode stays explicit via COPILOT_OFFLINE=true). Microsoft Execution Containers (MXC) went GA on Windows 11 for policy-enforced agent sandboxing, already supported by Codex, GitHub Copilot, OpenClaw, Replit, LM Studio and others, and llama.cpp support was added to Windows ML.