Hugging Face Releases LoRA Adapter for Qwen3.8-27B Coding Agent
Start here
Hugging Face's harness-hub team released Qwen3.8-27B-pi-adapted, an Apache-2.0 rank-32 LoRA adapter trained with AsyncGRPO reinforcement learning to run inside the pi coding agent. The team reports that with thinking off, it lifts SWE-bench Pro (hard 51) from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%, matching the base model's thinking-on accuracy at roughly 4.5x fewer tokens; on tasks both models solved, median turns on SWE-bench Pro fell from 59 to 23. These figures are self-reported on small evaluation sets, and the adapter only works with Qwen3.8-27B. The team also corrected an earlier 29.4% baseline cited in its tweet.
Generated from reports · updated 25 min ago
Report timeline
Follow the reports to see the event from different sides.
- MarkTechPostHugging Face Releases Qwen3.8-27B-pi-adapted LoRA, Lifting SWE-bench Pro From 39.2% to 64.7%
Hugging Face's harness-hub team released Qwen3.8-27B-pi-adapted, an Apache-2.0 rank-32 LoRA adapter trained with AsyncGRPO reinforcement learning to run inside the pi coding agent. With thinking off, it raises SWE-bench Pro (hard 51) from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%, matching the base model's thinking-on accuracy at roughly 4.5x fewer tokens. On tasks both models solved, median turns fell from 59 to 23 on SWE-bench Pro; all results are self-reported on small eval sets and the adapter only works with Qwen3.8-27B.
Heat for this event
Not enough continuous observations yet to draw a trend.