A team of researchers has published a full, reproducible recipe for turning a big open base model into a sharper one, and they show all their work.
The paper, posted to arXiv on September 25, 2026 (arXiv:2609.29421, https://arxiv.org/abs/2609.29421), describes Rufus-Air, a post-training pipeline built on GLM-4.5-Air-Base, a 106-billion-parameter model that activates 12 billion parameters per token. The recipe runs eight stages in order: supervised fine-tuning, then reinforcement learning for reasoning, coding, and instruction-following, then agent training for general tasks, coding, and search, and finally reinforcement learning from human feedback. The authors say every stage relies on open-source components and public data, with no new human annotation and no proprietary model used to generate training examples. They report that the result beats GLM-4.5-Air's official post-trained release and holds its own against other open models of similar size.
Post-training is usually the part labs keep closed even when the base model is open, because fine-tuning and reinforcement learning are where most real performance gains happen. Publishing the exact stage order, reward design, and data sources removes a lot of guesswork for anyone trying to replicate strong instruction-following and agentic behavior without a large annotation budget.
It is not a new model so much as a documented shortcut, useful if you already trust GLM-4.5-Air's base weights and less useful if you do not.