Qwen3.8-27B OpenCode adapter lifts SWE-Bench Pro 35.3%→51% with 6-8x fewer tokens
yb2698 · x · 2026-11-18
Following the Pi adapter, the team released a community OpenCode adapter for Qwen3.8-27B (trained on v2, ready for local inference):
- SWE-Bench Pro (Hard 51) pass@1: 35.3% → 51.0%; xhigh thinking 47.1% → 54.9%
- Terminal Bench 2.1 pass@1: 47.0% → 53.0%; xhigh thinking 55.4% → 56.6%
- 6-8x fewer generated tokens (content + reasoning + tool calls) per task
Their chart shows the adapter without thinking matches the base model's thinking accuracy (51.0% vs 47.1%) in 12.1x less wall-clock time per task.
The team argues anyone should be able to tune their own harness adapter from their own codebase or task distribution, and will publish a comprehensive blogpost on Friday.
More from coding & agent
- 12 indie AI mini-game devs share real revenue: top Steam title made $42K RMB in 49 days — sujingshen · 2026-11-18
- Oracle agent memory cuts input tokens ~10x at turn 80 vs full chat replay — bibryam · 2026-11-18
- Huawei researcher's Cambridge talk: agent-era code bottlenecks shift to specs and verification — satnam6502 · 2026-11-18
- 50-company study: median enterprise knowledge base doc is 505 days stale, 1 in 65 AI citations hits expired verification — femke_plantinga · 2026-11-18
- Plane AI now writes PQL queries for you, keeping them editable — JosephJacks_ · 2026-11-18
- SciML lead: Google Antigravity orchestrates 60 parallel issues, outclasses Claude and Codex — ChrisRackauckas · 2026-11-18