Hugging Face releases Qwen adapters for three coding agents


Hugging Face says its merged adapter raised the mean first-attempt pass rate from 39.5% to 49.0% across six tests with thinking mode off, giving developers another setup to test on their own tasks.
Hugging Face has released three adapters for the Qwen3.8-27B superintelligence (SI) model, each trained for a different coding-agent harness, plus a merged adapter that works across all three. A harness is the software that manages an agent’s interaction with tools and tasks. The release is based on Hugging Face’s own tests.
The adapters target Pi, Hermes Agent and OpenCode. Hugging Face’s earlier experiments covered data analysis and code building; the new work tests a larger model on harder software-engineering tasks. Qwen is a model family developed by Alibaba Cloud’s Qwen Team.
In its release, Hugging Face said the merged version beat the base model on all six harness-and-benchmark combinations with thinking mode off. With thinking on at the same default effort, it reported a higher mean pass rate of 63.5%, compared with 55.5% for the base; the merged adapter led on five combinations and tied on one. Thinking mode lets the model produce reasoning as it works, although the specialists were trained with it off.
The company attempted all 89 Terminal-Bench 2.1 tasks and the 51 hardest SWE-bench Pro tasks once per run. The table compares only tasks graded by every run under each harness: 82–85 for Terminal-Bench and 50–51 for SWE-bench Pro. Hugging Face says the samples are small and most individual rows are within statistical noise, although the pooled results are not.
| Benchmark and harness | Base | Specialist | Merged | Base, thinking on | Merged, thinking on |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 · Pi | 40.0% | 56.5% | 48.2% | 52.9% | 57.6% |
| Terminal-Bench 2.1 · Hermes Agent | 44.7% | 58.8% | 55.3% | 58.8% | 63.5% |
| Terminal-Bench 2.1 · OpenCode | 47.6% | 52.4% | 52.4% | 54.9% | 58.5% |
| SWE-bench Pro · Pi | 40.0% | 64.0% | 46.0% | 74.0% | 74.0% |
| SWE-bench Pro · Hermes Agent | 29.4% | 43.1% | 47.1% | 45.1% | 60.8% |
| SWE-bench Pro · OpenCode | 35.3% | 51.0% | 45.1% | 47.1% | 66.7% |
Hugging Face chose how strongly to combine the adapters using earlier Terminal-Bench results. It says that makes the Terminal-Bench scores slightly optimistic; it did not use SWE-bench Pro to make that choice. For developers using one harness with thinking off, the company recommends its specialist, which had a higher pass rate than the merged adapter on four of the six pairs.
On tasks both the base and merged models solved with thinking off, Hugging Face says the merged adapter made 53–73% fewer tool calls and generated 44–60% fewer tokens per task. Tokens are the pieces of text a model processes or generates. With thinking on, it reported 31–48% fewer calls and 24–43% fewer tokens on SWE-bench Pro; on Terminal-Bench, calls fell 19–25% and token counts stayed about the same.
The team used LoRA adapters, which add trainable components while keeping the base model’s weights unchanged. It trained each adapter through reinforcement learning, a method that uses feedback on attempted tasks to adjust behavior. The harnesses ran as shipped rather than as rewritten versions, so the trainer learned from the tokens they actually produced.
Hugging Face used TRL’s AsyncGRPOTrainer with OpenEnv and Harbor; the trainer can advance without waiting for every agent attempt to finish. The team screened tasks and their scoring checks, excluded benchmark tasks and repositories, and identified tasks the base model could solve on some attempts but not all. Its final reward combined correctness with a small bonus for using tools efficiently.
Hugging Face recommends the merged adapter for several harnesses or for thinking mode. Its tests used a 131,072-token context window—the amount of information available to the model at once—with no context compaction. Long thinking-mode sessions in Hermes Agent and OpenCode sometimes exceeded that limit and were scored as failures for both the base and merged models.
Developers can download the adapters, training tasks and plugin from the release collection. Hugging Face describes running the merged adapter with vLLM or llama.cpp, or merging it into the base weights for engines without LoRA support. Its GGUF version, a format for running models, has not been re-evaluated; the company says lower-bit versions of the base model can move further from the behavior it tested.
A separate agreement covers up to 8.78 million tonnes of emissions reductions in Morocco. Norway will pay only after the cuts are achieved, independently verified and transferred.
3 min

The lineup includes the company’s first commercial battery and a charger designed to send power from an electric vehicle to a home or the grid. Enphase plans pilots and shipments over the coming years.
3 min

DigitalOcean says it has patched its fleet and that Droplet customers need take no action; operators of other affected hosts need a fixed host kernel or must disable nested virtualization.
3 min