Arete-27B: Nous fine-tunes Qwen3.8-27B to strip state framing

An open model's values ship in its weights, and Nous Research just showed how thin that layer can be. Nous released Arete-27B, a fine-tune of Alibaba's Qwen3.8-27B that targets the base model's refusals and state-aligned answers on politically sensitive topics. Per Nous's own eval, answers that adopted the official line fell from 22% to 0.9%. The training budget was about 52.8 million tokens. For anyone running open models in production, that number is the story.

What actually happened

Nous published the Arete-27B write-up with weights on Hugging Face under Apache 2.0. The key details:

• Base model. Qwen3.8-27B, which Nous picked as the strongest local model and one that showed the behavior it wanted to fix.
• Method. Supervised fine-tuning only. No RL, no preference tuning, no architecture changes. Two epochs, 1,202 steps, 64k context. The teacher model was Inkling, chosen partly because an open teacher avoids distillation terms-of-use problems.
• The eval. 803 prompts across seven categories, including data privacy, surveillance and free markets. Each prompt carries 2–4 required facts from a dated source. Nous published the eval, the judge prompt and the scoring scripts.
• Results. Fact score rose from 2.32 to 2.98 out of 5. Propaganda answers fell from 22% to 0.9%. Nous reports no capability loss: 65.60% vs 65.77% on its internal Hermes Bench, and 70% vs 67% on Terminal-Bench 2.1 in its own harness.

Read the caveats too. Nous built the eval, picked the teacher and graded with an LLM judge (Grok 4.7). The thinking-on result is provisional, from two of five runs. Nous itself says it didn't fix every behavior tied to the base model's viewpoint.

Why open-model values matter for your business

If you self-host an open model, you inherit whatever its trainers put in. Usually that shows up on topics your support bot will never touch. But Nous cites outside reports that a model's built-in viewpoint can leak into tasks like code generation. You don't want to find that in production.

The flip side matters just as much. If 53 million tokens can remove a behavior, a similar budget can add one. A "fine-tuned" checkpoint from a stranger on Hugging Face is a black box until you test it.

So treat every open model the same way:

1. Write a small eval for your use case. Fifty real prompts with known-good answers beats any leaderboard.
2. Run it on the base model and any fine-tune before you switch. Check facts, refusals and tone.
3. Pin the exact weights you tested, by repo and revision. Don't pull "latest."

We said the same about the base Qwen3.8-27B release: the license is the easy part.

Sources: Nous Research, Hugging Face model card.