Z.ai has announced GLM-5.3, a model currently available only in its coding plan. Its arrival on the API is announced for soon, and its open weights are to be published on Hugging Face in two weeks. According to the company, the model surpasses Moonshot AI’s Kimi K3 on several benchmarks and Claude Fable 5 or GPT-5.6-Sol on some of them.
- Z.ai lance GLM-5.3, disponible pour l’instant uniquement dans son offre de codage.
- Le modèle repose sur une phase de post-entraînement fortement étendue, sans nouveau pré-entraînement.
- L’auteur rejette la distillation et attribue l’avance des laboratoires chinois à des sorties plus rapides et à la pression des classements.
- Z.ai restreint l’accès pour limiter les risques de cybersécurité, mais l’auteur juge ces mesures insuffisantes.
In its announcement, the company states that GLM-5.3 is built on the same base as GLM-5.2, with a significantly extended post-training phase. The blog post opens with this sentence: “Evolving post-training is all we did for GLM-5.3.” The model has about 750 billion parameters, which, according to the author of the Interconnects blog, is one-third of Kimi K3. The latter, which relies more on pre-training, is described as a “feat” in this area.
To explain how Chinese labs manage to stay on par with American models, the author dismisses distillation as a major factor, despite a recent research paper on reasoning trace extraction. He instead puts forward several elements. The first relates to release timelines: Z.ai could publish its models in a matter of days, whereas OpenAI and Anthropic often take months before a public release. According to the author, these American companies likely have more powerful internal models, but the testing time before publication allows Chinese labs to optimize their scores.
The author also mentions the pressure from public rankings, which, in his view, influence the company’s stock price and its ability to raise capital. He further believes that GLM-5.3 is probably more specialized than Claude Fable or GPT Sol, as Z.ai’s models are not multimodal. According to Bloomberg, cited by the blog, Z.ai is said to have reached $1 billion in ARR, largely thanks to its on-premise deployment business. Finally, the author highlights the company’s close ties with Tsinghua University.
A Model with Cyber Capabilities
Z.ai acknowledges that GLM-5.3 is its “most capable model to date for cybersecurity tasks,” with announced improvements in vulnerability discovery and exploit analysis. The company sees it as an aid for defenders, but also as a dual-use risk. It has therefore opted for a staged release: selected security partners will first evaluate the model in a controlled setting, before broader access and the publication of the full weights. Z.ai also states it monitors inference on its platforms using a query classifier and reasoning chain monitoring.
The author considers these measures useful, but limited, given that open-weight models of comparable capabilities exist or are arriving. He concludes that “this type of security barely matters” in this context, and argues for a direction “at an industrial scale, driven by public authorities or industry coalitions, to prepare immediately for this transition.”
