Arete-27B

Arete-27B is Nous Research’s supervised fine-tune of Qwen3.8-27B (named for Greek arete, excellence/virtue). It aims to keep the Chinese open base’s capability while reducing propaganda-framed or evasive answers on politically sensitive prompts.

Key facts:

• ~52.8M tokens, 64k context, 2 epochs, 1,202 steps; SFT only — no RL / preference opt;
• Teacher: Inkling; weights and train/eval data public;
• Page: https://nousresearch.com/arete ; HF: https://huggingface.co/NousResearch/Arete-27B

Why it matters

Open weights let labs re-align models to chosen values. Arete-27B ships a downloadable recipe plus a public eval, claiming little regression on general/agentic benches Nous ran.

Caveats

• “Western values” and propaganda-rate figures are Nous’s own eval + judge — independent replication is still thin;
• Nous does not claim all related behaviors are fixed;
• Disambiguate from philosophy/other Arete projects — use NousResearch/Arete-27B.

1. Read https://nousresearch.com/arete ;
2. Pull weights from https://huggingface.co/NousResearch/Arete-27B (Apache 2.0);
3. Re-run the public 803-prompt eval / arete-data-pairs yourself;
4. Audit safety and bias before production use.