GLM-5.3's agentic jump came with zero architecture changes. Terminal-Bench 3.0… · @harden_369741 · TwKit

@harden_369741

GLM-5.3's agentic jump came with zero architecture changes. Terminal-Bench 3.0 went from 4.6 to 28.3, all of it post-training environments, per http://Z.ai's own numbers. then a five-person team says they spent about $300k post-training the open Flash version and beat Claude Fable 5 on 4 of 6 knowledge-work benchmarks, $0.13 a task against $8.19. I used to refresh model release pages like iPhone launches. now I read the post-training notes first. the weights are the commodity, the environments are the moat. the catch nobody quotes: Andon Labs found the gains flatten on tasks http://Z.ai never trained for. post-training compounds, but only where you point it. so is the next frontier lab even a pre-training lab? #LLM #AI #OpenWeights

曝光 17 · 评论 0 · 点赞 0 · 书签 0