Z.ai has announced GLM-5.3. The model is currently available only in the coding plan, with the API coming soon and open weights arriving on Hugging Face in two weeks.

The jump in scores is striking. On many benchmarks the model surpasses Moonshot AI's Kimi K3; on some it passes Claude Fable 5 or GPT-5.6 Sol. That places it more or less at the frontier of agentic coding benchmarks.

The real number: 750 billion

What makes the result remarkable is not the score but the scale at which it was reached. GLM-5.3 does this with roughly 750 billion parameters — a third of Kimi K3.

Z.ai's blog post explains it directly, opening with a bold sentence: "Scaling post-training is all we did for GLM-5.3." The model uses the same base model as GLM-5.2; the difference comes entirely from substantially extended post-training.

At the risk of a broad oversimplification: Z.ai's strength appears to lie in post-training, while Kimi is more of a pretraining masterpiece.

Why the distillation answer falls short

The release revived a familiar question: how does China keep up so well? How can such a small model match the leading public American models? Are these results real?

The common answer is distillation — training a smaller model on a larger one's outputs. Nathan Lambert, who wrote the analysis, does not deem it the major factor.

His simplest explanation is different: Z.ai is very good at what it does. And it is worth recalling that this team has been working on this line of models longer than almost anyone in the industry.

A decade-long line

The history of the GLM family makes that concrete:

  • 2019: Zhipu AI founded
  • March 2021: GLM, released by THUDM, Tsinghua University's data mining group
  • August 2022: GLM-130B, the scaled version
  • March 2023: ChatGLM, the first chat version
  • June and October 2023: ChatGLM2 and ChatGLM3
  • January 2024: GLM-4; open-weight GLM-4-9B followed in June
  • February 2026: GLM-5, the latest major generation

GLM-5.2, released in June, mattered on its own terms. Weeks after release the author still heard from AI researchers using it — for its speed (some deploy it on internal clusters to run faster than public offerings) and its simplicity.

The uncomfortable part

The most candid passage is where the author describes his own denial: he admits thinking "how do they keep doing this? Surely the models aren't as good as they look."

His sharper observation follows: American companies hold a commanding resource lead, yet cannot seem to pull away in capabilities. That disconnect between resource gap and outcome gap is one of the least discussed issues in the industry.