An eight-stage training pipeline for GLM-4.5-Air is published in full

single source· 1 articles · confidence: medium · first seen 2026-09-23 20:00 UTC

What this means for you

Nothing to act on unless you post-train your own model. The claim worth reading is that stage order should follow reward reliability rather than capability, and that infrastructure and engineering choices are part of the recipe rather than an implementation detail. If you run a fine-tuning pipeline, that ordering heuristic is the thing to test.

Rufus-Air, posted to arXiv, documents an eight-stage pipeline for turning GLM-4.5-Air-Base — an open 106B-parameter model, 12B active per request — into a chat and agent model. Post-training is the stage after pretraining that turns a text predictor into an instruction-follower. The stages run from supervised fine-tuning through reinforcement learning for reasoning, coding and instruction-following, then general, coding and search agents, then RLHF (tuning against human preference scores). The authors report using public data throughout, with no new human annotation and no in-house teacher model to distil from, and say the result beats the official GLM-4.5-Air release. No benchmark numbers, harness or evaluation date are given.

Models in this story

Key facts

  • ·Rufus-Air is a post-training recipe built on GLM-4.5-Air-Base, given as 106B parameters with 12B active per token. source
  • ·The pipeline has eight stages in order: SFT, reasoning RL, coding RL, instruction-following RL, general agent, coding agent, search agent, RLHF. source
  • ·The authors report no new human annotation and no in-house distillation teacher, building on open-source components and public data. source
  • ·The paper claims the result improves on the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models; no benchmark figures are given. source
  • ·The paper is arXiv 2609.29421, listed on 23 September 2026. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire