GLM-5.3

What it is

GLM-5.3 is an open-weight large language model from Zhipu (also known as Z.ai), released on 14 August 2026. The weights, including those of a smaller GLM-5.3-Flash variant, were posted to Hugging Face on 25 August 2026^[2][3]. The release uses a custom GLM-5.3 license rather than a standard open-source licence, and the model card lists 753B parameters in BF16, F8_E4M3 and F32 tensor types^[1].

A fixed base, a retrained finish

GLM-5.3 uses the same base model as GLM-5.2, and Zhipu said every gain came from post-training^[1][2]. That claim is the most interesting thing about the release. Pre-training is the long, expensive stage in which a model reads very large amounts of text and learns general patterns of language and code. Post-training is everything after that: further training on curated examples, instruction-following data and feedback signals, which shapes how the model behaves on specific tasks. By holding the base model fixed, Zhipu turned the release into a demonstration of how much can still be gained at the second stage alone, without a new pre-training run.

Two settings

The model card describes two practical controls. A `reasoning_effort` setting takes the values low, high or max, and when the parameter is not passed it defaults to max; the card recommends keeping that default when reproducing benchmark numbers^[1]. Separately, `clear_thinking` defaults to false, and the card recommends passing `clear_thinking=true` in chat scenarios^[1].

Results Zhipu reported

Zhipu reported a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench, and open-source state of the art on Terminal Bench 3.0 and Agents' Last Exam^[1]. On individual benchmarks it gave the following pairs, GLM-5.3 first and GLM-5.2 second: Terminal Bench 2.1, 88.2 against 81.0; DeepSWE v1.1, 66.9 against 46.2; SWE-Marathon, 42.5 against 19.4; and HLE with tools, 62.5 against 54.7^[1].

These benchmarks all measure work rather than recall. Terminal Bench asks a model to complete tasks through a command line; DeepSWE and SWE-Marathon test software engineering over longer stretches of work; the HLE figure was recorded with tools available, meaning the model could call outside programs rather than answer from memory alone. The pattern across the numbers is that the largest jumps come on the long-horizon coding and agentic tasks, where a model has to plan several steps and recover from its own mistakes, while the smaller jump is on terminal use, which was already strong.

Cybersecurity

Zhipu also reported cybersecurity scores: CyberGym 84.5 against 77.2 for GLM-5.2, and more than double GLM-5.2 on exploitation benchmarks, with ExploitBench at 54.4 against 24.4^[1]. Capability of this kind matters in both directions. It is useful for finding and fixing vulnerabilities, and it is also the kind of capability that governments and safety teams track when a model's weights are downloadable by anyone.

Outside assessments

In September 2026, CAISI, the centre at NIST that evaluates AI systems, found GLM-5.3 the most advanced open-weight model in cybersecurity, and Anthropic rated its cyber capability as similar to Claude Mythos Preview^[2]. Those are third-party judgements rather than vendor claims, and they concern a model that anyone can download.

Where it sits

The release is a data point in a broader shift: as base models converge, post-training becomes the main place where capability is added, and a fixed base plus a better finishing stage can move agentic and security benchmarks substantially. GLM-5.3 also arrives with a license of its own rather than an established open-source one, and with independent evaluations of its cyber capability already on record — two things that shape how such a model can be used and discussed.