A reinforcement-learning adapter for Qwen3.8-27B is reporting higher coding benchmark scores while making fewer tool calls. The results come from the pi-specific adapterβs model card on Hugging Face: developer-reported tests, not independent replication. Itβs an upgrade trained for one coding agent, not a universal tune-up for every setup. Source
The scoresβand the fine print
The model card describes a LoRA adapter trained to work behind the pi coding agent on software-engineering and terminal tasks. Its results table reports these comparisons against the base model, also running under pi: Source
The SWE-bench figures cover only that 51-task selection. For Terminal-Bench, the table lists 86 tasks, while the methods section says all 89. The card doesnβt explain the difference. Source
Across all evaluated tasks, the card reports 55% fewer tool calls on the SWE-bench selection and 62% fewer on Terminal-Bench. Fewer calls are encouraging; they arenβt a stopwatch. Source
The evaluations used one attempt per task at temperature 0.8, with thinking disabled for both models. Each task had a 100-agent-step limit, with one hour allowed for Terminal-Bench and two hours for SWE-bench Pro. Source
The team picked the model saved at training step 160 using scores on SWE-bench Pro and an 18-task Terminal-Bench subset. Those tasks did double duty: they helped select the model version as well as measure its performance. On the 68 Terminal-Bench tasks not used to make that choice, the card reports an increase from 39.7% to 55.9%. Source
Donβt borrow OpenCodeβs scoreboard
Separately, TeksEdgeβs November 19 post describes an approximately 300 MB adapter optimized for OpenCode, reporting a rise from 35.3% to 51.0% on SWE-bench Proβs hard-51 selection. Those are TeksEdgeβs claims about a different adapterβnot the pi adapterβs size or results. The pi model card does not verify that separate OpenCode release. Source Source
Comments