Z.ai ships GLM-5.3 open-weight: same base as GLM-5.2, bigger post-training

Z.ai published GLM-5.3 as open-weight on Hugging Face on August 28. The model reuses the same base as GLM-5.2, and every gain comes from post-training, according to the model card. The company frames the result around three threads: stronger coding, longer-horizon agentic tasks, and cyber capability that it says emerged as post-training scaled.

What actually shipped

GLM-5.3 is a Mixture-of-Experts model with roughly 753 billion parameters in total; the config shows eight experts active per token. The released checkpoint includes an FP8 variant, and the card lists English and Chinese support. The model is steered with a reasoning_effort parameter that takes low, high or max, defaulting to max.

The license is custom. Hugging Face lists it under the name "glm-5.3" in the "other" category, which is not Apache 2.0 or MIT. If you plan to build a product on top of it, that license is the first thing to read. The card also points to a technical report, "GLM-5: from Vibe Coding to Agentic Engineering," on arXiv.

What the numbers say

The vendor claims are specific. Z.ai calls GLM-5.3 "the most capable open-weights model for coding," says it improves 50% over GLM-5.2 on its in-house Z.ai Code Bench, and claims open-source SOTA on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam. The card publishes the full comparison table against GLM-5.2, Kimi K3, DeepSeek-V4 Pro-0813, Qwen3.8-Max, Opus 4.8, Fable 5 and GPT-5.6 Sol, so the claims can be checked.

Against GLM-5.2, the lead is consistent and large: 88.2 to 81.0 on Terminal Bench 2.1, 66.9 to 46.2 on DeepSWE, and 54.4 to 24.4 on ExploitBench. The open-weights reading holds on CyberGym, where GLM-5.3 scores 84.5, the top of the table. On several coding rows the closed models still lead: GPT-5.6 Sol is ahead on Terminal Bench 3.0 (34.6 to 28.3) and on DeepSWE (72.7 to 66.9), and Fable 5 leads ProgramBench. The card's footnotes describe evaluation runs with up to 1M-token context, which is where the agentic task rows come from.

The cyber story is the bigger step, and Z.ai is unusually direct about it: as post-training scaled, cyber capability developed faster than the company expected, and the gains are largest further up the exploitation chain, where GLM-5.3 more than doubles GLM-5.2. Whether that capability is a safety concern or a defensive-security asset depends on who is running the model, and the card does not take that argument.

Two days earlier: Ox Alpha

The release lands two days after TechCrunch reported that Z.ai is the lab behind Ox Alpha, a mystery model that had been circulating without a public owner. Read together, GLM-5.3 is the public, open-weight counterpart to that stealth release, from the same lab and days apart.

If you're self-hosting

The card lists a wide serving surface: SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, plus Ascend NPU support through vLLM-Ascend and xLLM. Two practical notes if you run it. Keep the default reasoning_effort=max if you want to reproduce the leaderboard numbers, and pass clear_thinking=true in chat scenarios, as the card recommends. If you are evaluating it for a product, start with the license; the weights being downloadable does not answer what you are allowed to ship.