AIINEWS

AIINEWS

Global intelligence on artificial intelligence and computing

LoRA adapter helps Qwen3.8-27B solve more coding tasks in the pi agent

AIINEWS Editorial Board•frontier-models
19 November, 2026•2 min read
LoRA adapter helps Qwen3.8-27B solve more coding tasks in the pi agent

Executive Briefing

“Hugging Face’s harness-hub team has released a LoRA adapter that tunes Qwen3.8-27B for the...”

Key Takeaways

  • 1.The harness-hub team released Qwen3.8-27B-pi-adapted, a LoRA adapter trained with reinforcement learning for the pi coding agent rather than general chat.
  • 2.The team’s self-reported evaluations used pi and one attempt per task. With thinking disabled, success rose from 39.2% to 64.7% on 51 hard SWE-bench Pro tasks and from 39.5% to 55.8% on 86 Terminal-Bench 2.1 tasks.
  • 3.For tasks both versions solved, median interaction turns fell from 59 to 23 on SWE-bench Pro and from 18 to 8 on Terminal-Bench 2.1.
  • 4.On SWE-bench Pro, the base model also scored 64.7% with thinking enabled. The reported figures say the adapter generated about 4.5 times fewer tokens than the base model in that mode.

Hugging Face’s harness-hub team has released a LoRA adapter that tunes Qwen3.8-27B for the pi coding agent. In the team’s reported results, its success rate on a SWE-bench Pro task set rose from 39.2% to 64.7% with thinking disabled.

Key Developments & Primary Facts

The harness-hub team released Qwen3.8-27B-pi-adapted, a LoRA adapter trained with reinforcement learning for the pi coding agent rather than general chat.

The team’s self-reported evaluations used pi and one attempt per task. With thinking disabled, success rose from 39.2% to 64.7% on 51 hard SWE-bench Pro tasks and from 39.5% to 55.8% on 86 Terminal-Bench 2.1 tasks.

For tasks both versions solved, median interaction turns fell from 59 to 23 on SWE-bench Pro and from 18 to 8 on Terminal-Bench 2.1.

On SWE-bench Pro, the base model also scored 64.7% with thinking enabled. The reported figures say the adapter generated about 4.5 times fewer tokens than the base model in that mode.

Strategic Analysis & Implications

The notable result is not only a higher solve rate with thinking disabled. Within pi, the adapter reached the reported SWE-bench Pro score of the thinking-enabled base model while using fewer tokens. That could lower serving costs, but token counts alone do not establish actual savings.

Because the adapter was trained for pi, these results are most relevant to deployments using that agent. Gains measured in pi do not establish that it improves other coding agents.

Outlook & Factors to Watch

The self-reported evaluations use small task sets. The supplied information contains no independent replication.

The supplied information does not disclose inference VRAM requirements, so the minimum hardware needed to deploy the adapter remains unclear.

Independently verified, edited, and published by AIINEWS Editorial Board
3-tier verified
Editorial standards
Share this intelligence report:
Expanded Perspectives & Intelligence

Related Dispatches & Deep Dives

Curated 8 specialized reports

Explore AIINEWS Intelligence Vault

Over 2,000 in-depth dispatches structured across 18 specialized tech & AI categories.

Browse all 18 categories