

Arete-27B
#11 in Open-Source-SpraakmodelleNous Research · 27b · siet 6. November 2026 (Ankündigung); offizieller Launch laut einer Quelle am 7. November 2026 · 8× · tolest 10. Nov. 2026
58
Momentum
Arete-27B is a supervised fine-tune by Nous Research built on Alibaba's Qwen3.8-27B. According to the provider, it aims to reduce state-aligned answers on politically sensitive topics. Weights, training data and evaluation are publicly available on Hugging Face.
Momentum-Verloop
18.08.16.11.
Features
| Key Benchmark (%) | Propaganda rate 22% → 0.9% (own 803-prompt eval); Hermes Bench 65.60% vs. 65.77% base |
| Context Window (Tokens) | Trained at 64K; base model native 262,144 tokens (not separately stated for Arete) |
| License | Apache 2.0 |
| Multimodality | Text-only training; vision encoder from base model frozen |
| Platform | Hugging Face (open weights, self-hosted) |
| Release Date | November 6, 2026 (announcement) |