Arete-27B: Nous open-sources a re-aligned Qwen3.8, data and evals

Apache-2.0 weights, a 93k-pair training set, and an 803-prompt eval all ship — plus TwIL-8B, a local reasoner tuned to spend fewer tokens thinking.

Nowline NOV 10 1:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A re-aligned open 27B, with its whole recipe

    Nous Research SFT-tuned Alibaba's Apache-2.0 Qwen3.8-27B on ~52.8M tokens to cut outputs its own rubric scores as CCP 'propaganda,' then open-sourced the lot: BF16 weights, the 93.4k-pair training set, and the full 803-prompt eval. Nous reports the propaganda rate dropping 22% to 0.9% with capability flat (Hermes-Bench 65.8% vs 65.6%), but those numbers come from its own judge (Grok 4.7) and aren't independently replicated yet — so the real value is a complete, Apache-2.0 template for re-aligning an open model and measuring the shift.

  • An 8B tuned to think in fewer tokens

    webAI's TwIL-8B post-trains IBM's Apache-2.0 Granite-4.2-8B with a final length-aware RL stage that rewards shorter reasoning, shipping Balanced (cheaper) and Verbose (max-accuracy) modes; webAI reports 98.2 on MATH-500 and 81.6 on LiveCodeBench. It runs locally in ~16 GiB at bf16 (less when quantized) but ships under a non-commercial license, so it's built for local tinkering and research rather than anything you'd put in production.