Language Models · Directory rank #6
Arete-27B
Nous retrained Qwen’s values.
Watch the reel
Arete-27B is Nous Research’s post-training of Qwen3.8-27B, retrained to change how it answers sensitive questions.
Fact score went up from 2.32 to 2.98, and propaganda answers dropped to under 1 percent. It is the same model, retrained values.
Training was supervised fine-tuning only — about 52.8 million tokens, two epochs, teacher model Inkling — and general capability stayed flat on the benchmarks the authors ran.
Key facts
- Nous Research post-trained Qwen3.8-27B
- Retrained to change how it answers sensitive questions
- Fact score up from 2.32 to 2.98
- Propaganda answers down to under 1 percent
- Evaluated on 803 prompts, judge Grok 4.7 blind to the model
- General capability stayed flat (Hermes Bench 65.77% vs 65.60%)
PikyAI has not tested it live. This page is a model card overview from the PikyAI reel.
PikyAI reel on YouTube. Watch on Instagram
How to use it
- Hugging Face par model page kholo.
- Model download karo aur apne setup par chalao.
- Sensitive sawalon ke jawab check karo — values retrained hain.
Pros and cons
Pros
- Fact score up — 2.32 se 2.98.
- Propaganda answers 1 percent se neeche.
- Qwen3.8-27B ki base par bana hai.
Cons
- Licence verify nahi hui — page par koi licence claim nahi.
- Authors khud kehte hain ke sab behaviors fix nahi hue — fully fixed nahi hai.
- PikyAI ne live test nahi kiya.
Alternative
- Qwen3.8-27B — Qwen3.8-27B asal base model hai — Arete usi ki retrained version hai.
How this rank works
Directory rank is set by bid, not by editorial opinion. New listings start at $5, a higher bid outranks a lower one, and taking #1 costs at least $5 more than the current top bid. Paid ranks always carry a Sponsored label, and payments are final and non-refundable. This listing sits at the $5 floor as a PikyAI house listing, not a paid rank.