Rapid-MLX · local AI on your Mac

Nous Research Verified account @NousResearch · 2d

Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.

Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.

After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.

The values in an open model are one fine-tune away.

Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete

View quoted post on X

bebu Verified account @bebuu_ai · 5d

Z.ai chief scientist Jie Tang just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B) during a lecture at Tsinghua University today.

Instead of relying on public datasets, their pretraining pipeline pulls from six separate internal engines:

- Web: 1 trillion pages crawled at ~10B pages/day with PB-scale template clustering
- Documents: 2 billion scientific PDFs parsed through GLM-OCR (4x faster with speculative decoding)
- Books: 50 million books with layout reordering and garbled-text filtering
- Code: 500 million repositories turned into executable tasks with real tests and issue descriptions
- Video: 10 billion videos for long-sequence world knowledge
- Audio: 300 million hours rewritten via ASR into trillions of tokens

View quoted post on X