AINEWS
Z.ai, a Chinese AI lab, says it caught up to America's biggest AI on coding tests not by building a larger model, but by training the one it had longer.
On August 14, 2026, Z.ai, a Chinese AI lab formerly Zhipu and spun out of Tsinghua, released GLM-5.3: a roughly 750-billion-parameter coding model that scored near the top of the agentic coding tests the company chose to highlight. That is about a third the size of Moonshot AI's Kimi K3, its closest Chinese rival, and a fraction of the parameter counts behind leading US frontier systems, the top tier of public AI from labs like OpenAI and Anthropic.
What changed is not the model. Z.ai's own blog post opens with: "Scaling post-training is all we did for GLM-5.3." GLM-5.3 shares the same base as its predecessor, GLM-5.2 from June 22, 2026. The difference is in post-training, the stage that comes after pretraining, where engineers fine-tune a finished base model on narrower, higher-quality feedback: code that runs, answers that satisfy a rubric, agent trajectories completed end to end. Interconnects analyst Nathan Lambert reports that post-training is "substantially extended" for GLM-5.3 relative to GLM-5.2. Same parameter count, more rounds of feedback.
This is the piece of the AI race the scoreboard usually misses. The conventional story is that frontier position comes from scale: more parameters, more compute, more data. Z.ai's claim with GLM-5.3 is that you can hold the base model fixed and squeeze out competitive performance by spending that effort in post-training instead.
The benchmark table in the announcement is Z.ai-reported, not independently verified. Interconnects' summary flags results that beat Moonshot's Kimi K3 plus a partial edge over what the post names as "Claude Fable 5" and "GPT-5.6-Sol" on some tests. Weights and a model card are headed to Hugging Face in roughly two weeks. Until then, the post-training scaling claim is exactly the kind of result that should be checked against the released model card.
Z.ai's track record on post-training is not a fresh bet. The GLM family starts at THUDM in Tsinghua in March 2021, with GLM-130B in August 2022, ChatGLM in March 2023, and GLM-4 in January 2024. Zhipu, the predecessor company, was founded in 2019 and rebranded as Z.ai along the way. GLM-5 landed on February 11, 2026, and GLM-5.2 on June 22. Interconnects notes GLM-5.2 was popular with researchers for inference speed and for shipping without rollbacks. The pattern is a lab that has shipped repeatedly and learned what to fix in post-training.
The interesting comparison when the Hugging Face weights land will be what GLM-5.3 did differently in post-training relative to GLM-5.2, and whether the coding gains hold on tasks that are not coding.