Alibaba's Qwen3.8-27B runs Opus 4.6-class coding benchmarks on a laptop

Alibaba released the open weights for Qwen3.8-27B, a dense 27-billion-parameter model licensed under Apache 2.0, alongside the larger 2.4-trillion-parameter Qwen3.8 mixture-of-experts model. The 27B version is small enough to run on a well-specced MacBook Pro or Mac Studio, and Unsloth's 4-bit quantized build needs only 17GB of GPU memory, within reach of a single consumer card.

On Alibaba's own benchmark comparisons, Qwen3.8-27B beats Anthropic's Opus 4.6 running at its Max setting on LiveCodeBench v6 (90.3 vs. 88.8) and on instruction-following via IFBench (79.5 vs. 62.5). Opus 4.6 Max still leads on Terminal Bench 2.1 (78.2 vs. 73.0) and on repo-level code generation, where longer-context reasoning across a real codebase matters more than isolated benchmark tasks.

For a team already paying per-token for a hosted coding agent, a 17GB local model that wins on some benchmarks and trails on others is a real tradeoff, not a clean replacement. It's a credible option for offline work, latency-sensitive tasks, or cutting API spend on lower-stakes completions, while repo-scale refactors likely still favor a hosted model with deeper context handling. Community MLX conversions for Apple silicon are already available, at roughly 16.1GB for the 4-bit build and 29.5GB for 8-bit.