Qwen3.6-27B codes locally at 50 tok/s

This is not a real benchmark, but it is the kind of field report that matters: local coding models are crossing from novelty into “could actually sit in a workflow.”

• –Qwen’s own model card claims 27B parameters, native 262K context, and strong coding-agent scores, including 77.2 on SWE-bench Verified and 59.3 on Terminal-Bench 2.0.
• –The interesting signal is not raw speed alone; it is that a single high-end consumer GPU can run a long-context quantized coding model at interactive latency.
• –Claude Code and Opus-class systems still have the product polish, tool reliability, and reasoning edge, but open weights are closing the gap on basic planning and repo navigation.
• –The caveat is large: one Reddit test does not prove day-to-day reliability, especially for tool calls, long sessions, and messy production codebases.