Arete-27B demonstrates targeted Qwen retuning

A targeted retune changes Qwen’s answer tendencies

Nous Research’s release of Arete-27B shows that a targeted fine-tune can materially alter how an existing open-weight Qwen model handles politically sensitive questions, without rebuilding its architecture or producing a meaningful capability regression on the tests Nous ran. The developer’s internal evaluation found that answers classified as adopting the Chinese government’s official position fell from 22% for Qwen3.8-27B to 0.9% for Arete-27B, a reduction of 21.1 percentage points.

The result demonstrates the effect of additional training rather than proving that the resulting model is politically neutral. The evaluation measures adoption of one specific official position; it does not establish that other forms of bias were eliminated or that the model is generally more truthful.

Fine-tuning preserved most measured capability

Arete-27B is a 27-billion-parameter dense model built from Alibaba’s Qwen3.8-27B through supervised fine-tuning. Nous used about 52.8 million tokens generated by the open Inkling teacher model over two training epochs. The company reported no reinforcement learning, preference optimization or architecture changes, while the vision encoder remained frozen because the training data was text-only.

The model is distributed under the Apache 2.0 license. MarkTechPost reported that its community Q4_K_M quantization is 17.20 GB, sized for 24 GB consumer GPUs, providing a practical route to local deployment.

On Nous’s Hermes Bench, Arete-27B scored 65.60%, compared with 65.77% for the base model. Nous characterized that difference as noise. The model card also reported a Terminal-Bench-2.1 result of 70%, versus 67% for the base model, but declined to claim an improvement because the test configurations differed.

Factual coverage improved, but remained limited

Arete-27B’s fact score rose from 2.32 to 2.98 out of five in Nous’s evaluation of 803 English-language questions, with five runs per model. Answers receiving a zero fact score fell from 359 to 216. The largest category change was in freedom of expression, where the fact score increased from 0.89 to 2.44 and the state-position adoption rate declined from 60% to 1%.

These figures indicate better reproduction of required facts under Nous’s rubric, not broad factual accuracy. MarkTechPost reported that 216 of 803 answers still received a zero fact score, while the source evaluation explicitly cautions that a low adoption rate does not mean the model can accurately explain nearly all facts.

Inspectable evaluation still needs external testing

Nous published the trained weights, training data, evaluation prompts and grading scripts, allowing researchers to inspect or rerun the comparison. The evaluation used Grok 4.7 as an automated judge under fixed settings: reasoning mode disabled, temperature set to zero and maximum output length of 32,000 tokens. The 803 prompts covered seven political and social topics, with each prompt containing two to four required facts.

Because the grading was performed by an AI model using criteria written by the developer, these results have not been independently reproduced by a third party. The immediate verification point is whether independent evaluations using different judges and scoring rules reproduce the shift in state-position adoption without materially degrading general capability.

The reviewed reporting disclosed a model release and downloadable research artifacts, but did not establish customer procurement, commercial deployment or an open sourcing process.

Industry and product implications