Skip to content
NewA calmer, faster ZeroHour

LoRA adapter tunes Qwen3.8-27B for the OpenCode coding agent

Original title: harness-hub/Qwen3.8-27B-opencode-adapted — new model trending #26 on Hugging Face

Hugging Face trending modelsby harness-hub ()9 min read
Part of a story covered by 4 sourcesRadixArk releases DaisyRose-27B with company-reported Terminal-Bench gainsMerged summary & timeline
AI summary · x-ai/grok-4.7Summaries by Vani AI

The publisher reports 51.0% on SWE-bench Pro hard 51 with thinking off, above the base model’s high-effort score.

harness-hub released Qwen3.8-27B-opencode-adapted, a reinforcement-learning LoRA for Qwen/Qwen3.8-27B meant to drive the OpenCode agent on software-engineering and terminal tasks. Under its own OpenCode setup, thinking off, it claims 51.0% versus 35.3% for the base model on SWE-bench Pro hard 51, and higher Terminal-Bench 2.1 scores with fewer tokens. GGUF weights and serving notes for vLLM, transformers and llama.cpp are linked. The full training recipe is deferred to a later blog post.

Unconfirmed: All benchmark figures are the publisher’s own OpenCode-harness results and are not comparable with the base-model card.

Key facts
  • Checkpoint step 150: SWE-bench Pro 51.0% versus 35.3%, 11 newly solved and 3 lost.
  • On 66 held-out Terminal-Bench tasks the card cites +13.6 points, p = 0.035.
  • Median turns on jointly solved SWE-bench tasks are 13 versus 48 for the base model.
  • Evaluation uses one attempt, a 100-step budget, temperature 0.8 and text only.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

9 min read at huggingface.co

image-text-to-text · 302 downloads · 111 likes · license apache-2. 0

Hugging Face trending models

By harness-hub · huggingface.co

Read the full article at Hugging Face trending models

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.