GLM-5.3 shows how Chinese labs keep pace without distillation

Z.ai released GLM-5.3, an open-weight model of roughly 750 billion parameters, and Nathan Lambert uses it to answer a question that keeps coming up: how do Chinese labs stay so close to the frontier? The easy answer has been distillation, quietly training on the outputs of American models. Lambert argues that story is mostly wrong, and GLM-5.3 is his evidence. The model is an extension of GLM-5.2 through heavy post-training, and Z.ai is blunt about it: "Scaling post-training is all we did for GLM-5.3." It beats Moonshot's Kimi K3 on agentic coding benchmarks and matches several closed American models.

Instead of distillation, Lambert lists concrete reasons. Chinese labs ship days after a model is ready rather than months, which lets them keep tuning against public benchmarks. They optimize for those benchmark scores harder than US labs do. They buy access to the same reinforcement learning environments and data through commercial markets. And a lab like Z.ai draws deep engineering talent from Tsinghua. None of these are shortcuts; they are ordinary execution done fast.

There is a tradeoff worth holding onto. Models tuned this way can be narrower than they look, strong in coding or security but less general than a single headline benchmark suggests.

Why it matters

If you choose open-weight models, GLM-5.3 puts frontier-level coding at open-weight prices within reach. But Lambert's point doubles as a warning: benchmark-topping Chinese models may be tuned narrowly, so test on your own tasks before assuming the score transfers.