GLM 5.3 Explained

GLM 5.3 is Z.AI's current flagship model. Released on 2026-08-17, it is designed for coding, long-horizon tasks, and cybersecurity work. According to the model card on Hugging Face, it has 753 billion parameters and supports a maximum context of 1 million tokens. It uses the same base model as GLM 5.2, with Z.AI attributing the improvements to post-training rather than a new base.

The weights are publicly available, but unlike GLM 5.2, they are not released under the MIT licence. Instead, GLM 5.3 uses a bespoke licence named after the model. GLM-5.3-Flash, published seven days apart, uses the standard MIT licence.

What GLM 5.3 Is

GLM 5.3 is a large language model (LLM) published by Z.AI, the international brand of the Chinese lab Zhipu. Z.AI released it on 2026-08-17 and pointed it at three jobs: writing code, running long-horizon tasks, and cybersecurity work.

Z.AI serves GLM 5.3 through its own application programming interface (API) under the model id glm-5.3, so an integration has to carry that exact string to reach GLM 5.3. The web chat at chat.z.ai runs on the smaller GLM-5.3-Flash instead, which means the chat window and the API do not answer with the same model.

GLM 5.3 and GLM-5.3-Flash are also the only two models the GLM Coding Plan serves. A subscription request naming GLM 5.2 or GLM 5.1 is routed to GLM 5.3 automatically, with no error raised.

• Released 2026-08-17 by Z.AI, aimed at coding, long-horizon tasks and cybersecurity
• Callable API model ids are glm-5.3 and glm-5.3-flash, so an integration reaches this model only with the exact string
• Weights are published on Hugging Face as zai-org/GLM-5.3, which puts self-hosting in reach of anyone whose legal review clears the licence
• The GLM Coding Plan serves GLM 5.3 and GLM-5.3-Flash only, so a team there still asking for GLM 5.2 is being answered by GLM 5.3

GLM 5.3 Shares GLM 5.2's Base Model

Z.AI says GLM 5.3 uses the same base model as GLM 5.2 and attributes the improvement to post-training instead of a new pre-training run. That leaves the size of the upgrade genuinely open.

A new base model means a fresh set of learned weights. Post-training reshapes behaviour on top of weights that already exist. So every figure Z.AI publishes for GLM 5.3 describes what post-training did, and no outside lab has reproduced any of it.

Anyone already running GLM 5.2 can measure the difference directly, because GLM 5.2 stays callable on the API at its published rates. One workload, run twice, settles the question for that workload better than a benchmark table does.

• Same base model as GLM 5.2, per Z.AI, so the whole gap between the two releases rests on post-training
• No new pre-training run is claimed, which is why the 5.3 label describes changed behaviour rather than a rebuilt model
• No independent reproduction of the gains has been published, so the improvement is unmeasured outside Z.AI
• GLM 5.2 remains callable at the same published rates, which makes an A/B test on one real workload cheap to run

GLM 5.3 Parameters and Context Window

GLM 5.3 carries 753 billion parameters and a 1-million-token maximum context, both published on its Hugging Face model card. The million-token ceiling lets a whole repository, or a long agent run with its full history, sit inside a single request.

GLM-5.3-Flash is a separate and smaller model at 320 billion parameters. Z.AI puts that one behind its web chat, so the size difference shows up for anyone comparing chat answers against API answers.

A 753-billion-parameter model does not fit on a single consumer graphics card, which makes cluster hardware the practical floor for self-hosting rather than a workstation. Running GLM locally covers what that takes.

• GLM 5.3: 753 billion parameters, per the Hugging Face model card zai-org/GLM-5.3
• GLM-5.3-Flash: 320 billion parameters, and the model Z.AI runs behind chat.z.ai
• Maximum context: 1 million tokens, enough to hold a full codebase or a long agent transcript in one call
• Both models are downloadable, so the real barrier to self-hosting is hardware and the licence, not access to the files

Routing a Task to GLM 5.3 or GLM-5.3-Flash

GLM 5.3 and GLM-5.3-Flash differ in parameter count, published price and weights licence, so routing between them is not only a speed decision.

GLM 5.3 costs about 19 times as much per million input tokens as GLM-5.3-Flash, and about 18 times as much on output. Any task GLM-5.3-Flash handles correctly costs about a nineteenth as much, so GLM 5.3 earns its price only where the cheaper model gets the answer wrong.

GLM-5.3-Flash also ships under plain MIT. A team that wants weights it can host without a bespoke licence review has only one of the two models to choose from, whatever the benchmark tables say.

GLM 5.3 is a poor fit for two groups. A company that cannot get a non-standard weights licence through legal review should stay on GLM-5.3-Flash or on GLM 5.2's MIT weights. So should anyone who needs an independently reproduced benchmark before deploying, because none exists for this release yet.

Two developments would change that reading. If Z.AI published the licence terms plainly on the model card, or if Artificial Analysis posted a reproduced score for GLM 5.3, the case for it would rest on checkable evidence instead of on Z.AI's own reporting.

The GLM 5.3 Licence Is Not MIT

GLM 5.3's weights ship under a bespoke licence named after the release itself, glm-5.3. GLM 5.2's weights were MIT. GLM-5.3-Flash, seven days apart, shipped under plain MIT, so two models from the same family and the same month carry different terms.

Published weights and open source are not the same thing. Anyone can download GLM 5.3 from Hugging Face. What a company may then do with the files is set by the licence text in that repository, not by the fact that the download works.

The New Stack reports that the licence requires companies above a large aggregate-revenue threshold to pass a Z.AI security review before hosting GLM 5.3 commercially. Z.AI does not state that threshold on the model card. No dollar figure circulating for it should be treated as confirmed, and the LICENSE file in the GLM 5.3 repository is the text to read before any commercial deployment.

• GLM 5.3 weights: a bespoke licence named glm-5.3, so an earlier MIT review does not transfer to this release
• GLM 5.2 weights: MIT, which is why a team that cleared 5.2 has to send 5.3 back through legal
• GLM-5.3-Flash weights: plain MIT, published seven days apart from GLM 5.3
• The revenue threshold The New Stack reports is not on the model card, so the licence file is the only text worth planning against

Z.AI's Benchmark Claims for GLM 5.3

Every GLM 5.3 benchmark figure in circulation comes from Z.AI, and none has been reproduced by an independent lab. For a procurement file, a number with one source is a claim rather than a measurement.

Z.AI reports a 50% improvement over GLM 5.2 on Z.AI Code Bench, a benchmark it runs internally. On Terminal-Bench 3.0 it reports 28.3, up from 4.6. On CyberGym it reports 84.5%, against 83.8% for Anthropic's Mythos 5 and 83.6% for OpenAI's GPT-5.6 Sol.

Z.AI also reports 2,436 confirmed vulnerabilities found across 269 open-source projects, 1,097 of them critical or high severity. Artificial Analysis publishes model evaluations run outside the labs themselves and is the place to check whether any of these figures survive. The same treatment of the previous release sits in the GLM 5.2 benchmarks guide.

• Z.AI Code Bench: a 50% gain over GLM 5.2, on a benchmark Z.AI runs itself
• Terminal-Bench 3.0: 28.3, up from 4.6, reported by Z.AI
• CyberGym: 84.5% for GLM 5.3, 83.8% for Mythos 5, 83.6% for GPT-5.6 Sol, all reported by Z.AI
• 2,436 confirmed vulnerabilities across 269 open-source projects, 1,097 critical or high severity, reported by Z.AI
• None of these has been reproduced independently, so a decision resting on them rests on a single party's own numbers

What GLM 5.3 Costs

Z.AI publishes GLM 5.3 at $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens. GLM-5.3-Flash is published at $0.075, $0.015 and $0.25 on the same three lines, which is cheap without being free.

The GLM Coding Plan is a separate subscription route, and Z.AI's documentation publishes only a starting price of 18 USD per month for it rather than a per-tier table. Full rates, credit tiers and the quota windows sit in the Z.AI pricing guide and on Z.AI's own pricing page.

What GLM 5.3 Changes for Teams Already Running GLM

The licence change reaches procurement first. A legal review that cleared GLM 5.2's MIT weights does not carry over, because GLM 5.3's terms are new text written for a single release. For companies weighing a Chinese-hosted model at all, the security-risk guide covers the wider review.

The second change is quieter and lands on engineering. Because the GLM Coding Plan serves only GLM 5.3 and GLM-5.3-Flash, a request naming GLM 5.2 or GLM 5.1 is answered by GLM 5.3 with no error and no warning.

A team that benchmarked GLM 5.2 on that plan, then left the model string untouched, is being served GLM 5.3 rather than the model it benchmarked. Requests naming GLM 4.7 route to GLM-5.3-Flash under the same rule. Neither routing rule applies to the direct API, where GLM 5.2 and GLM 5.1 stay callable at their published rates.

That split between the two surfaces is easy to miss, and worth checking before assuming a fixed model sits behind either. Background on Z.AI and its wider model line is in the Z.AI explainer. Before committing to GLM 5.3, read the LICENSE file in its Hugging Face repository and re-run one real workload against both GLM 5.3 and GLM-5.3-Flash.

• A GLM 5.2 legal review does not cover GLM 5.3, so the licence goes back through counsel before any self-hosted rollout
• Coding Plan requests for GLM 5.2 or GLM 5.1 are served by GLM 5.3 silently, which invalidates a benchmark run against the older string
• Coding Plan requests for GLM 4.7 are served by GLM-5.3-Flash, so a cost model built on GLM 4.7 rates stops matching the bill
• The direct API still serves GLM 5.2 and GLM 5.1, which is the route to take when a pinned model version is a requirement

What you need to run GLM 5.3 yourself

GLM 5.3 is a frontier-scale Mixture-of-Experts model, so "running it yourself" is a real infrastructure decision — not something a single laptop or gaming GPU can do. Match the path below to how seriously you need to self-host. For most teams the API or rented GPUs are the right answer; buying hardware only pays off at steady, high volume or when your data can never leave your walls.

Once GLM 5.3 is running, the fastest way to put it to work day to day is inside Cursor — point it at the model through OpenRouter as a custom model. And if you would rather run a model on one affordable box, see Best mini PCs for local AI and Local AI hardware calculator.

NVIDIA RTX 6000 Ada (48GB)

Power users and small clusters that want cards they own

View on Amazon →