ToolsNov 18, 2026 · 20:371 source

OpenCode Adapter Raises Qwen3.8-27B Coding Score to 51% and Cuts Tokens 6-8x

From marketbrief.now

Image: marketbrief.now
AI;DR

A new LoRA adapter for Alibaba's Qwen3.8-27B model optimized for the OpenCode v2 harness increased the model's pass@1 score on the SWE-Bench Pro (Hard 51 subset) from 35.3% to 51.0%. The adapter also raised the model's Terminal-Bench 2.1 pass@1 score from 47.0% to 53.0% and reduced the number of generated tokens per task by 6 to 8 times.

The release follows a similar adapter for the Pi coding agent released Nov. 17, which more than doubled the model's SWE-bench Pro score. The developers aim to adapt models to specific software interfaces, or harnesses, rather than adjusting the interfaces to fit the models. A full reproducible recipe for creating these harness-specific adapters will be published this Friday.

Read the sourceSign in to suggest

Sources

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today