NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Zhipu's GLM-5 Keeps 744 Billion Parameters but Activates Only 40 Billion

Zhipu released GLM-5 with 744 billion total parameters and 40 billion active per token, open weights under MIT, and benchmark results it says lead open models on coding and agent tasks.

2026-11-21 · 997 words · NeuroAI
Zhipu's GLM-5 Keeps 744 Billion Parameters but Activates Only 40 Billion

A coding agent needed to fix a bug in a codebase it had never seen. The old answer was to give it a bigger model and a bigger bill. Zhipu's answer this month was to give it a bigger model that mostly sleeps.

The company, also known as Z.ai, published GLM-5 around November 19, 2026, with a technical report and release blog titled "From Vibe Coding to Agentic Engineering." The framing is a claim about how software gets built: not an assistant that completes a function, but a system that reads a repository, finds the fault, writes a patch, runs the tests and iterates.

The numbers, read from the source

The official release states that GLM-5 scales from GLM-4.5's 355 billion parameters, with 32 billion active, to 744 billion parameters with 40 billion active. Pre-training data rose from 23 trillion to 28.5 trillion tokens. The model integrates DeepSeek Sparse Attention, or DSA, to reduce deployment cost while preserving long-context ability.

Those are the figures to use. They come from the company's own technical announcement, not from an aggregator.

A parameter total of 744 billion with 40 billion active per token is a mixture-of-experts design. The capacity is stored; only a fraction is computed for any given token. That is how a model can grow by a factor of two in total size while active compute rises by only a quarter.

About the other numbers you may have seen

Several round-up posts in Chinese media described GLM-5 as having 745 billion parameters, 78 hidden layers, 256 experts and a 202K context window. Those figures should not be repeated without checking.

The official announcement says 744 billion parameters. It does not give a layer count or an expert count in the release blog. The "202K" figure appears to be a misreading of a footnote: the report says HLE-with-tools evaluations used a maximum context of 202,752 tokens. That is a setting for one benchmark, not the model's context window. The SWE-bench suite, in the same footnote, is run with a 200K context window.

This kind of drift matters. A reader choosing hardware for deployment needs the real context limit, not a number rounded up from an evaluation parameter. When a round-up and a technical report disagree, the technical report wins.

What the model is claimed to do

Zhipu reports a SWE-bench Verified score of 77.8 and a Terminal-Bench 2.0 score of 56.2, or 60.7 on a verified version of that benchmark that fixes ambiguous instructions. It reports 62.0 on BrowseComp, rising to 75.9 with context management, and 89.7 on the tau-squared bench for tool-using agents.

The most unusual number comes from Vending Bench 2, a test in which a model runs a simulated vending machine business over a one-year horizon. Zhipu says GLM-5 finished with a final account balance of 4,432 dollars, ahead of other open models and close to Claude Opus 4.5's reported 4,967 dollars. The runs were conducted independently by Andon Labs, which gives that particular result more weight than the company's internal evaluations.

Post-training used an asynchronous reinforcement-learning system called slime, developed with Tsinghua University, which Zhipu says raises training throughput and allows finer-grained iteration. The company describes a shift from "vibe coding," where a human prompts and debugs, toward agents that close their own loops.

Read the benchmark table carefully

Almost every score in the announcement is measured by Zhipu. The footnote specifies temperatures, token limits, judge models and context settings. Those details are good practice, but they also show how much a score depends on the harness.

SWE-bench Verified is a public benchmark, but the version of the model, the sampling temperature and the prompt are chosen by the party running it. No third-party reproduction of the 77.8 figure was available at the time of writing. "Best among open models" is the company's characterisation, not an independent certification.

The Vending Bench result is the cleaner signal precisely because an outside organisation ran it. A model that keeps a small business solvent for a simulated year is doing something different from answering exam questions, and it is a better proxy for the agentic tasks the company is targeting.

Weights, licences and chips

GLM-5's weights are published on Hugging Face and ModelScope under the MIT licence. The model is available through api.z.ai and BigModel.cn, supports Claude Code and OpenClaw, and can be tried free on Z.ai. Local deployment is documented for vLLM and SGLang.

The chip list is notable: Zhipu says the model can be deployed on hardware from Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame and Hygon, through kernel optimisation and quantisation. That is a long way from saying the model was trained without Nvidia hardware, a claim that has appeared in third-party commentary but not in the official release. Compatibility at inference time is a verifiable statement; origin of training hardware is not established here.

Honest limitations

All specifications, benchmark results, licence terms and chip compatibility statements come from Zhipu's official GLM-5 release blog and its footnotes. I did not run the model or inspect the weights. Benchmark figures other than Vending Bench 2 are vendor-measured.

The 745B, 78-layer, 256-expert and 202K figures circulating in Chinese media are not used here, for the reasons given above. If a later official document confirms an expert count, that detail can be added; it is not asserted now.

I did not cover GLM-5.3 or GLM-5.3-Flash, which appear in round-ups as separate commercial models, because they are different releases. Nothing here is a recommendation to buy or deploy a particular model.

What readers can do now

  • If you are planning deployment, use the official context and active-parameter figures. Hardware sizing based on a round-up number will be wrong.
  • If you build coding agents, try the model on your own repository. A public benchmark score says little about your test suite and your failure modes.
  • If you care about supply chains, check which chip platform actually works for you. Stated compatibility and acceptable throughput are two different claims.

Related coverage

Cite this article

NeuroAI (脑机网), "Zhipu's GLM-5 Keeps 744 Billion Parameters but Activates Only 40 Billion," 21 November 2026, https://neuroai.site/a/na-model-glm5-745b

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories