Hugging Face announces Qwen3.8-27B adapters for three agent harnesses
Developers using pi, Hermes Agent or OpenCode can keep the same base model and choose between a specialist add-on and one that works across all three.
Hugging Face announced four adapters for Qwen3.8-27B on November 20 for developers using pi, Hermes Agent or OpenCode. The set includes a specialist for each harness and a combined version, Qwen3.8-27B-super-adapter, that works across all three.
The adapters let developers change how Qwen3.8-27B responds to prompts and uses tools while keeping the same base model and harness. The harness manages the prompts, tool calls and context the model receives. Each adapter uses low-rank adaptation, or LoRA, to apply a small set of trained weight updates while leaving the base weights unchanged.
Hugging Face says it trained the specialist adapters with reinforcement learning inside each unmodified harness. The company says it then combined the three updates at half strength without further training.
In tests of the combined adapter, Hugging Face reported that the average first-attempt pass rate rose from 39.5% with unadapted Qwen3.8-27B to 49.0% with the adapter when thinking mode was off. With thinking mode on at the same xhigh reasoning effort for both versions, Hugging Face reported an increase from 55.5% to 63.5%.
The averages span six pairings of the three harnesses with SWE-bench Pro and Terminal-Bench 2.1. The SWE-bench Pro sample came from its 51 hardest tasks, and the comparison within each harness used only tasks graded in every run. Hugging Face used one attempt per task and a 100-step limit in text-only runs.
Hugging Face cautioned that the Terminal-Bench 2.1 results may be optimistic because it used that benchmark to choose the merge weight. The announcement cites no independent replication of these results. The model card says Hugging Face has yet to repeat the tests with the combined adapter’s GGUF build.
Developers can download all four adapters in GGUF format. The combined adapter’s model card documents serving it with vLLM or loading its GGUF file alongside the base model in llama.cpp.
More news

Baidu details ERNIE Evolve automated research system

Snowflake makes Cortex Agent SDK generally available

Signal adds mobile controls to limit AI agents' access to chats
