Microsoft Foundry Local runs Qwen3.8-27B on-device for Copilot

Microsoft's Foundry Local now runs Alibaba's Qwen3.8-27B as a local agentic coding model, plugging it into GitHub Copilot through a Bring Your Own Key hook and serving it on ONNX Runtime. The November 5 post from the Foundry team reports Qwen3.8-27B scoring 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro, with 42.3 on NL2Repo-Bench and 42.2 on DeepSWE 1.1.

The floor is a single NVIDIA GPU with 24 GB of VRAM; 4-bit weight quantization shrinks the model to roughly 18 GB. Context runs up to 64K tokens on a 24 GB card and up to 262K tokens on 32 GB and above, with tests cited on an NVIDIA GeForce RTX 5090 and an NVIDIA DGX Spark with 128 GB of unified memory.

Under the hood the stack leans on DFlash2 speculative decoding, INT4 weight quantization, KV-cache quantization, prefix caching for concurrent requests, CUDA graph fusions tuned for NVIDIA SM121, and CPU offloading. Authors Baiju Meswani, Emma Ning and Tianlei Wu frame the pitch bluntly: "Local inference flips that trade-off: your code and your context never leave the machine in front of you."

Mac support, AMD GPU support, additional coding-optimized local models and a deeper GitHub Copilot integration sit on the roadmap. CLI version 0.13.0 or later is required. The release drops into a crowded coding-tools beat on our tracker.