Z.ai GLM-5.3

Z.ai's GLM-5.3 is a 753-billion-parameter mixture-of-experts model, with 40 billion parameters active per token, a one-million-token context window and text as its only input. It is the open-weights release this reading picked: its weights landed on 28 August 2026, inside the last thirty days, and it carries more dated evaluations than any other model whose weights shipped in that window.

We read four Z.ai pages (the launch post, the model card, the licence file and the API documentation), two independent sources (an evaluator's index page and an analysis of the release), two dated reports, one benchmark tracker and the two community sources a plain page fetch could open, the Hacker News thread from launch day and the repository's own discussions. The dated material runs from 14 August to 16 September 2026, and the pages themselves were read on 17 September 2026. Nothing was run, installed, prompted or tested.

Consensus

The evidence agrees on what GLM-5.3 is: a long-horizon coding and agentic model made by taking the GLM-5.2 base and spending more post-training on it, rather than a new pretraining run. Z.ai's launch post puts it plainly, and its model card and API documentation repeat it, that every gain over GLM-5.2 comes from post-training.

The one independent index we could read has it first among open-weights models. Artificial Analysis scores GLM-5.3 (max) 45 on Intelligence Index v4.3, ahead of Kimi K3 at 44, and calls it amongst the leading models in intelligence while noting that it is particularly expensive beside open-weights models of similar size, slower than average and very verbose. On Z.ai's own table it is first among open weights on Terminal-Bench 3.0, at 28.3 against 4.6 for GLM-5.2, on Agents' Last Exam at 28.5 and on CyberGym at 84.5. Both readings agree that coding is where the model moved, and they part company on how much of that reaches a reader who is not running agent harnesses.

Confidence is uneven. The official pages are strong: the model card, the licence file and the documentation were read directly. The independent side is one evaluator's index and one analysis of the release, both dated. Developer evidence is one long thread and a handful of repository discussions. The vendor's table is a claim, and no independent run of its sixteen benchmarks was readable on 17 September.

Recurring strengths

It is the open-weights leader on the independent index. Artificial Analysis places GLM-5.3 (max) first among the open-weights models it ranks, at roughly a fifth of Kimi K3's list price per million tokens. That is the index reading in this sample.

The coding gain is large, and it comes from post-training alone. Z.ai reports Terminal-Bench 3.0 rising from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9 and Agents' Last Exam from 23.8 to 28.5 against GLM-5.2 on the same base model. Its own model card is the source, and its own model card is a claim.

The licence stays permissive for almost everyone. The licence file grants use, copying, modification, fine-tuning, distribution, sublicensing and sale, and adds one condition: a licensee running a Model-as-a-Service business whose revenue passes ten billion dollars over any twelve consecutive months must pass Z.AI's security review before commercial use. For nearly every reader the file is MIT in everything but its name.

Recurring complaints

The weights arrived two weeks after the model did. The launch post said the weights would follow two weeks after launch, once safety evaluation and hardening were complete, and attributed the hold to cyber capability developing faster than expected. Reports date the weights to 28 August, with the repository's last modification at 15:22 UTC that day. One analysis writes down 25 August instead.

It is verbose, and it is priced against the closed frontier. Artificial Analysis reports about 210 million output tokens across its evaluation suite against a median near 120 million for open-weights models, and calls the model slower than average. Z.ai's own documentation explains part of it: reasoning cannot be switched off, and the default effort level is max.

Running the whole checkpoint is a rack, not a workstation. The repository ships fp8 weights of several hundred gigabytes, and the analysis we read puts a multi-GPU node as the realistic minimum. The repository's download count reached 94,403 by 2 September with 1,478 likes, which measures interest rather than use.

Where reviewers split

How long the hold was. One analysis dates the weights to 25 August, three days ahead of the deadline Z.ai had set itself. Reports date them to 28 August, and the repository's last modification is that day. Both are consistent with a staged push, since the weights and the licence file need not arrive together.

The independent index moved under it. One report carried Artificial Analysis's first reading, 60 on Intelligence Index v4.1.1 on 18 August, tying Kimi K3. A tracker records 44.9 for the same model under v4.3 on 16 September. This is not a fall in the model: the index itself was revised between the two readings, so the two numbers are not the same measurement.

Where the cyber lead stops. Z.ai's table is first on CyberGym, at 84.5 against the 83.8 and 83.6 it lists for Mythos 5 and GPT-5.6 Sol, which is detecting and locating flaws. The same table has ExploitBench at 54.4 against 78.0 for Mythos 5 and 76.5 for GPT-5.6 Sol, which is further up the exploitation chain. The analysis adds that the two-hour and six-hour ExploitGym budgets are not hours on a clock: they are renormalised on each model's own tokens per second.

The vendor's table against the index. Z.ai's sixteen rows put GLM-5.3 at the top of the open-weights field on several benchmarks. Artificial Analysis's composite, which runs its own suite on its own infrastructure, places the model 8.5 points behind Claude Fable 5.1 on a 16 September reading. The two are not comparable, and each is honest about what it measured.

Who it suits

You run coding agents over long tasks and want weights you can host. The licence grants fine-tuning and redistribution, and the model card documents serving in SGLang, vLLM, TokenSpeed and Transformers.

You pay per token and want the strongest open-weights option an independent index scores. Artificial Analysis ranks it first among them, and Z.ai's list price is unchanged from the previous generation.

You want to study how much of a capability jump comes from post-training. The base model is the one the previous generation used, and the vendor says so on the card.

Who should pass

You want frontier-class output from a machine in the corner. The fp8 repository runs to several hundred gigabytes and the reports assume a multi-GPU node, so the download is the cheap part.

You need images in or out. The documentation and the model card both say text only, which several rivals at this size do not.

You need predictable bills. Reasoning is always on, max effort is the default, and the independent evaluation measured about 210 million output tokens where open-weights models of that size median near 120 million.

This is a reading of published reviews, not medical, financial or legal advice. For a decision about your health, your money or your rights, a qualified professional is the right next step, and not a review.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it. Get the weekly verdict.

Sources

• Z.ai launch post, official, 14 August 2026.
• GLM-5.3 model card, official, read 17 September 2026.
• GLM-5.3 License, official, read 17 September 2026.
• GLM-5.3 API documentation, official, read 17 September 2026.
• Artificial Analysis on GLM-5.3 (max), editorial, read 17 September 2026.
• noze, the weights, the licence and the benchmark method, editorial, 2 September 2026.
• Unite.AI report on the first index reading, news, 18 August 2026.
• Mervin Praison on the weights release, news, 28 August 2026.
• AI Atlas model page and benchmark observations, other, observed 16 September 2026.
• Hacker News launch thread, forum, 14 August 2026.
• zai-org/GLM-5.3 discussions, forum, read 17 September 2026.

On 17 September 2026 we read four official Z.ai pages, two independent sources, two dated reports, one benchmark tracker and two community sources, published between 14 August and 16 September 2026, by web search and plain page fetch, and synthesised them with AI; no model was run, no library installed and nothing was tested.