Qwen 3.8 27B: Apache-2.0 vision coder that fits one GPU
Tops Hacker News as builders test it: single-GPU vision, 262K context, Opus-class coding claims — plus a reasoning default you'll want to turn down.
Copy markdown
Runs on the GPU you already have
Dense 27B, Apache 2.0, with native vision and a 262K-token context (1M via YaRN). A ~17GB Q4 GGUF fits a single RTX 3090/4090 or an M-series Mac — a self-hostable multimodal model you can wire into a project this weekend with no per-token bill.
The Opus-class scores are vendor-only
Qwen's card claims SWE-bench Pro 61.7 (above Opus 4.6 Max's 53.4), LiveCodeBench v6 90.3, and OSWorld-Verified 84.3%. No independent runs confirm the coding numbers yet — treat the headline as a claim to test on your own repo, not a settled result.
Turn the reasoning down first
Simon Willison found the xhigh default wildly overthinks: an SVG of a pelican on a bicycle burned 22,276 reasoning tokens over 21 minutes. Set reasoning_effort to low or off for routine work — quality holds and you stop paying for tangents.
Speed is the real limit
On an M5 Max MacBook and an NVIDIA DGX Spark it runs ~15-30 tokens/sec — fine for private, offline, or batch work, but short of hosted APIs at 180+ for tight interactive agent loops. Local here means capable, not fast.
Elsewhere: Claude Code learns GitLab MRs
Claude Code v2.1.233 adds GitLab merge-request review to the code-review plugin, plus faster self-hosted runner starts and a Windows credential-leak fix. GitLab teams now get the same automated review loop GitHub users already had.