Amazon team's open Rufus-Air recipe post-trains GLM-4.5-Air on under 400 H200s
JingfengY · x · 2026-10-15
A small research team at Amazon Search built Rufus-Air, a fully open and reproducible post-training recipe from scratch on GLM-4.5-Air-Base (106B-A12B), documented in a 48-page arXiv paper.
- Scale: Never more than 400 H200s, using only public data and open-source components — no new human annotation and no in-house distillation teacher.
- Eight-stage serial pipeline: SFT → Reasoning RL → Coding RL → Instruction-Following RL → General Agent → Coding Agent → Search Agent → RLHF, progressing from basic to advanced capabilities and from hard verifiable rewards to softer judge-based signals.
- Key findings: Diverse, high-quality SFT sets a strong capability floor; difficulty filtering keeps RL prompts in a productive learning range; reward reliability is a practical principle for stage ordering; infrastructure and engineering choices are part of the recipe itself.
- Results: Outperforms the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models.
The paper documents data, reward design, infrastructure, stage order, and stagewise results for full reproduction.
More from Research
- New paper builds a categorical theory unifying hundreds of PEFT methods, proving LoRA rank limits — aryaman2020 · 2026-10-15
- ActionOPD distills agents at action level, lifting Qwen3 3.3-6.9 points on OOD benchmarks — nics-efc · 2026-10-15
- Horizon Robotics' VITAL hits 90.8 PDMS with a single denoising step at 3.29x inference speed — HorizonRobotics · 2026-10-15
- KAIST's KV4-RL: INT4 KV cache quantization speeds RLVR rollouts 1.8x without accuracy loss — kaist-ai · 2026-10-15
- ASU's τ-Emotion-Bench shows LLM agents stay unreliable when users get emotional — Arizona-State-University · 2026-10-15
- Huawei's SkillMarker watermarks agent skills, detecting copied passages of 1,536 tokens — Alan Tang · 2026-10-15