Open Source for LLM Post-Training: What Can We Build?

A lightning talk given at Open Infra Night for Agent Builders, part of Open Source AI Week in San Francisco. It asks how far a small team can take a base model with open-source resources alone. Our answer is Rufus-Air, a post-training effort at Amazon Rufus led by Tuo Zhao: an open recipe on GLM-4.5-Air-Base with eight serial stages, from hard rewards to soft ones, built on slime, Megatron-LM, SGLang and Strands Agents. The talk then zooms into the harness stack behind its agentic stages, the small details that silently break agentic RL, and finally introduces tokin, an open-source gateway that makes any agent harness token-native…