GLM-5.3: How Chinese labs keep stride with the frontier
| Source: Interconnects (Nathan Lambert)
Tags: GLM-5.3, Z.ai, Zhipu AI, post-training, open-weights, agentic coding, Kimi K3, Chinese AI labs
Z.ai's GLM-5.3 matches frontier coding benchmarks — including beating Kimi K3 and Claude Fable 5 on some tests — using only 750B parameters (a third of Kimi K3), achieved entirely through post-training improvements on the GLM-5.2 base model.
Details
Z.ai announced GLM-5.3, a coding-focused model that sits at the frontier of agentic coding benchmarks despite having roughly 750B parameters — about a third of Moonshot AI's Kimi K3. The model outperforms Kimi K3 on many benchmarks and beats Claude Fable 5 or GPT-5.6-Sol on some, positioning it alongside the leading publicly available models. What makes GLM-5.3 notable is the simplicity of the recipe: 'Scaling post-training is all we did,' according to Z.ai. The model uses the same base as GLM-5.2 — a model that had already earned a following among researchers for its speed and clean behavior — but with substantially extended post-training. Nathan Lambert argues this is evidence of Z.ai's particular strength in post-training, contrasting with Kimi's pretraining-first approach. The popular explanation that Chinese lab performance is driven by distillation from frontier models is challenged here. Lambert attributes Z.ai's consistent performance to genuine deep expertise — the lab has been working on GLM since 2019, longer than almost any competitor. GLM-5.3 is currently in coding plans only, with API access and open weights via HuggingFace expected within two weeks.