Nous Research, a US startup known for its open-source agent Hermes, fine-tuned Qwen3.8-27B to remove Chinese state-aligned responses. Its test used 803 prompts across seven categories, and a judge model checked each answer against two to four facts drawn from dated sources. The unmodified model scored 2.32 out of 5 on factual accuracy, and 22 percent of its answers counted as propaganda.
Training used supervised fine-tuning without reinforcement learning or preference optimization. Arete-27B scored 2.98 on factual accuracy and 0.9 percent of its answers were propagandistic. The largest change came in the free speech category, where the propaganda rate fell from 60 to one percent.
On Nous's internal Hermes Bench, Arete-27B scored 65.60 percent against 65.77 percent for the base model. Nous does not count a rise on Terminal-Bench 2.1 from 67 to 70 percent as a real gain, because the test setups differed. The weights, test dataset and training data are available under an Apache 2.0 license on Hugging Face.
Why it matters if you build agents
Teams building agents on Chinese open-weight models may see state-aligned answers on political topics, and this result suggests that such bias can be removed with a relatively small supervised fine-tune. Because the weights and training data are published, the method and test set can be checked before relying on them. The benchmark figures are self-reported by Nous, so performance should be verified on your own workloads.
Get stories like this every weekday
One short email with the day's agent tooling news.