AI Trends

Trending story

Hugging Face Trains Qwen3.8-27B Framework Adapters

1 reports1 story updatesUpdated Nov 20 08:00

Understand this story

AI overview

On November 20, 2026, Hugging Face trained LoRA harness adapters for Qwen3.8-27B tailored to pi, Hermes Agent, and OpenCode, and merged the three specialized adapters into a super-adapter capable of running across different frameworks. Hugging Face said all three specialized adapters outperformed the base model in Terminal-Bench 2.1 and SWE-bench Pro evaluations within their respective frameworks. With thinking mode disabled, the super-adapter increased average pass@1 from 39.5% to 49.0% and reduced the number of tool calls needed to solve tasks collectively by 53% to 73%. Training used in-framework reinforcement learning, TRL AsyncGRPOTrainer, and 871 agent coding tasks. The related adapters, GGUF files, reinforcement-learning tasks, and safeagents plugin have been released publicly. There are currently no other substantiated developments.

AI-generated from public reports

Latest development · Report

Hugging Face Releases Qwen3.8-27B Harness Adapters, Improving Agentic Coding Performance and Efficiency

Developments

1 confirmed developments

  1. Hugging Face Trains LoRA Adapters for Different Agent Frameworks

    Source:Hugging Face

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 20

  1. Hugging FaceFirst-partySelected

    Hugging Face Releases Qwen3.8-27B Harness Adapters, Improving Agentic Coding Performance and Efficiency

    Hugging Face has released LoRA harness adapters for Qwen3.8-27B tailored to pi, Hermes Agent, and OpenCode, and merged the three into a super-adapter capable of running across different frameworks. Each of the three specialized adapters outperformed the base model in Terminal-Bench 2.1 and SWE-bench Pro evaluations within its respective framework. With thinking mode disabled, the super-adapter increased average pass@1 from 39.5% to 49.0%, while reducing the number of tool calls needed to solve tasks jointly by 53–73%. Training used in-framework RL, TRL AsyncGRPOTrainer, and 871 agentic coding tasks. The related adapters, GGUF files, RL tasks, and safeagents plugin have all been made publicly available.

Related stories

  1. Related storiesXiaomi MiMo Releases MiMo-V2.6 Pro and Flash
  2. Related storiesRadixArk Uses Miles to Train DaisyRose-27B
  3. Related storiesSGLang natively supports and powers the Beam production API
  4. Related storiesMeta releases open-weight model Muse Spark Open