In brief: Z.ai’s new model GLM-5.3, with around 750 billion parameters — a third the size of Kimi K3 — matches leading US models on agentic coding benchmarks, achieved solely through massively expanded post-training on the same base as GLM-5.2.
Z.ai has unveiled GLM-5.3, a model that outperforms Moonshot AI’s Kimi K3 on numerous agentic coding benchmarks and in some cases also beats Claude Fable 5 and GPT-5.6-Sol — with only around 750 billion parameters, a third the size of Kimi K3. For practitioners, this raises once again the question of how Chinese labs keep pace at the frontier of model development with significantly fewer resources.
GLM-5.3 is initially available only through Z.ai’s Coding Plan, but is expected to arrive via the API soon and as open weights on Hugging Face in two weeks. According to Z.ai’s blog post, GLM-5.3 is built on the same base model as GLM-5.2 but received substantially expanded post-training. The company puts it succinctly: “Scaling post-training is all we did for GLM-5.3.” Compared to Kimi, which is regarded more as a pretraining reference, Z.ai is thus positioning itself as a specialist in post-training methods.
The model series dates back to 2019, when Zhipu AI was founded. The original GLM (General Language Model) was released in March 2021 by THUDM, the data mining group at Tsinghua University, followed by GLM-130B in August 2022 and the ChatGLM versions in 2023. GLM-4, released in January 2024, marked the switch to today’s model naming, with the open GLM-4-9B variant added in June 2024. GLM-5 followed in February 2026, and GLM-5.2 in June of the same year. According to the author, some AI researchers continue to use GLM-5.2 internally because of its speed and ease of use, in some cases on their own clusters for faster inference.
Regarding the debate over why Chinese labs are catching up, the author cites several possible explanations. Distillation from frontier models is a common explanation but is not considered the main factor. The author mentions a recent study on simple methods for extracting reasoning traces from frontier models — a technique Chinese labs could be using at scale, and one that, according to the author, US providers have so far not consistently prevented. In its blog post, Z.ai itself describes an RL-dominated training regime with “more environments, more diverse tasks, and more compute to train on them” — infrastructure and algorithms for reinforcement learning environments, the author notes, cannot simply be distilled.
For practitioners evaluating models for agentic coding workloads, the relevant point is that GLM-5.3 delivers competitive benchmark results with a significantly smaller parameter count and is set to become available as open weights shortly. The question of benchmaxxing — that is, deliberate optimization for test sets rather than real-world generalization — remains unanswered in the original text and is flagged by the author as an open point of discussion.
Source: www.interconnects.ai · Published August 14, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.