Announcement
@yb2698 announces Qwen3.8-27B adapter for local OpenCode use
The model card reports similar accuracy to Qwen’s xhigh mode, using 6–8× fewer tokens with thinking off.
TLDR
According to harness-hub’s model card, the reinforcement-trained adapter lifts Qwen3.8-27B from 35.3% to 51.0% on SWE-bench Pro’s hard-51 subset with thinking off. Its pitch: less token-churning, not more thinking. The reported 6–8× token reduction is versus the base model’s xhigh mode, not its thinking-off runs. Results come from text-only OpenCode tests and aren’t directly comparable with Qwen’s published benchmarks.
Combined views
1.2K
3 Sources, first seen ago

Comments