---
format: "aidr-story-markdown/v1"
id: "e41f6931145e154887cb533f8c3e1c0617491908a5a799b4d4bf36a5001ac346"
canonical_url: "https://aidr.today/e41f6931?lang=en"
title: "Qwen3.8-27B SWE-Bench Pro Score Hits 63% With New Pi Harness Adapter"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-17T13:52:12.000Z"
category: "Tools"
topics: ["llm","devtools","open-source","coding"]
source_urls: ["https://marketbrief.now/ai/qwen38-27b-swe-bench-pro-score-hits-63percent-with-new-pi-harness-adapte-5cd41442","https://huggingnews.com/ai/qwen38-27b-swe-bench-pro-score-hits-63percent-with-new-pi-harness-adapte-5cd41442","https://huggingnews.com/ai/pi-team-to-release-reproducible-recipe-for-harness-adapters-after-qwen38-85415f89","https://marketbrief.now/ai/pi-team-to-release-reproducible-recipe-for-harness-adapters-after-qwen38-85415f89","https://huggingnews.com/ai/pi-harness-adapter-more-than-doubles-qwen38-27bs-swe-bench-pro-score-045548b4","https://marketbrief.now/ai/pi-harness-adapter-more-than-doubles-qwen38-27bs-swe-bench-pro-score-045548b4","https://www.marktechpost.com/2026/11/18/hugging-face-qwen3-8-27b-pi-adapted-lora/"]
summary: "A new LoRA adapter for Alibaba's Qwen3.8-27B model increased its pass@1 score on the SWE-Bench Pro (Hard 51 subset) from 29% to 63% when used with the Pi harness. The adapter also raised the model's Terminal-Bench 2.1 pass@1 score from 40% to 56% while reducing the number of tool calls per task by 55% to 62%. The team trained the adapter using reinforcement learning through TRL's AsyncGRPOTrainer and OpenEnv, employing real Pi sessions in a Hugging Face sandbox. The developers intend to release more harness-specific adapters in the coming days and will publish the full training recipe in a blog post this Friday."
---

# Qwen3\.8\-27B SWE\-Bench Pro Score Hits 63% With New Pi Harness Adapter

> [Open the canonical story](<https://aidr.today/e41f6931?lang=en>)

**Published:** 2026-11-17T13:52:12.000Z
**Category:** Tools
**Topics:** llm, devtools, open\-source, coding

## Summary

A new LoRA adapter for Alibaba's Qwen3\.8\-27B model increased its pass@1 score on the SWE\-Bench Pro \(Hard 51 subset\) from 29% to 63% when used with the Pi harness\. The adapter also raised the model's Terminal\-Bench 2\.1 pass@1 score from 40% to 56% while reducing the number of tool calls per task by 55% to 62%\. The team trained the adapter using reinforcement learning through TRL's AsyncGRPOTrainer and OpenEnv, employing real Pi sessions in a Hugging Face sandbox\. The developers intend to release more harness\-specific adapters in the coming days and will publish the full training recipe in a blog post this Friday\.

## Sources

- [Story source](<https://marketbrief.now/ai/qwen38-27b-swe-bench-pro-score-hits-63percent-with-new-pi-harness-adapte-5cd41442>)
- [Story source](<https://huggingnews.com/ai/qwen38-27b-swe-bench-pro-score-hits-63percent-with-new-pi-harness-adapte-5cd41442>)
- [Story source](<https://huggingnews.com/ai/pi-team-to-release-reproducible-recipe-for-harness-adapters-after-qwen38-85415f89>)
- [Story source](<https://marketbrief.now/ai/pi-team-to-release-reproducible-recipe-for-harness-adapters-after-qwen38-85415f89>)
- [Story source](<https://huggingnews.com/ai/pi-harness-adapter-more-than-doubles-qwen38-27bs-swe-bench-pro-score-045548b4>)
- [Story source](<https://marketbrief.now/ai/pi-harness-adapter-more-than-doubles-qwen38-27bs-swe-bench-pro-score-045548b4>)
- [Story source](<https://www.marktechpost.com/2026/11/18/hugging-face-qwen3-8-27b-pi-adapted-lora/>)

