Open coding agents, small LoRA adaptations and reinforcement learning

Hugging Face trains a Qwen3.8 coding-agent adapter to use fewer tokens

The pi-specific Apache-2.0 LoRA improves paired coding-agent scores in publisher tests, but is neither a standalone model nor an independent universal benchmark.

Open Model Weights published19 Nov 2026
Primary sourceHugging Face harness-hub — official Qwen3.8-27B pi-adapted model card
Source published2026-11-18 (new model-card update; exact initial timestamp not shown on card)
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original minimalist editorial illustration representing code in AI research; conceptual, not actual model output.
AI-generated editorial illustration · Not official model imagery.

A new Hugging Face harness-hub release suggests a way to improve coding agents without retraining an entire foundation model. The team has published an Apache-2.0-declared LoRA adapter for Qwen3.8-27B, trained specifically inside the pi coding-agent environment. Instead of building another full 27B checkpoint, the project changes how an established model performs software-engineering and terminal workflows through a compact set of adapter weights. Primary source ↗

According to its official card, the selected adapter came from reinforcement learning over verifiable software tasks. The training setup used AsyncGRPO, the Harbor and OpenEnv environments, pi as the harness and correctness rewards with a small tool-efficiency bonus. It selected a rank-32 LoRA checkpoint at training step 160. The approach is intentionally harness-specific: a change in agent tools, prompts or evaluation environment can change results.

The paired publisher evaluation reports SWE-bench Pro's 51 hard tasks rising from 39.2% solved by the base model to 64.7% with the adapter. On the 86 Terminal-Bench 2.1 tasks that both systems graded, the figures move from 39.5% to 55.8%. The team describes one attempt per task and 100 agent steps, with thinking disabled. Those constraints matter because the Qwen base model's headline scores in other configurations use different harnesses and budgets.

An efficiency finding may be more consequential than the pass-rate increase. The authors say their adapter with thinking off approximates the accuracy of the base model at a high reasoning setting while writing about 4.5 times fewer generated tokens per task. That does not mean every workload becomes 4.5 times cheaper: output lengths, pricing, inference infrastructure and completed-task rates are different quantities. Nor should the claim be confused with a model-wide, independent ranking.

The evidence has useful transparency and a caution. Hugging Face revised an earlier SWE-bench result after discovering that public-tunnel failures interrupted several baseline rollouts, then reran affected evaluations. The published paired numbers are therefore preferable to older screenshots or social posts. The project also states that its GGUF conversion has not been separately benchmarked, and users should not expect equivalent scores automatically from more aggressive quantization.

To try the release, an operator still needs the original Qwen3.8-27B base model, an appropriate LoRA-compatible runtime and the pi tool-calling template. The card gives examples for vLLM, a merged Transformers model and llama.cpp. The lesson for open-weight builders is specific: training in the real agent harness may improve verifiable task completion and token efficiency, but the adapter is not a new standalone foundation model.

OMW REGISTRY WATCH

What this changes for the evidence layer

This is an Apache-2.0-declared LoRA adapter requiring Qwen3.8-27B base weights, not a new 27B independent foundation checkpoint. Verify separate adapter files, exact repo SHA and base license before any standalone model record. The card corrected an earlier test result after tunnel failures; cite the current paired results only.

PRIMARY SOURCE

Hugging Face harness-hub — official Qwen3.8-27B pi-adapted model card

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗