GLM-5.3: How Chinese labs keep stride with the frontier

Nathan Lambert published this on interconnects on 2026-08-14, the day z-ai announced GLM-5.3. At launch the model was available only through Z.ai's coding plan. API access was promised soon, and open weights on Hugging Face in two weeks. The benchmark jump was large. On many benchmarks GLM-5.3 beat Moonshot's Kimi K3, and on some it beat Claude Fable 5 and GPT-5.6-Sol. That puts it roughly at the frontier of agentic coding with about 750B parameters, a third of kimi-k3's size.

Z.ai's blog opens with one sentence: "Scaling post-training is all we did for GLM-5.3." It is the same base model as GLM-5.2 with much longer post-training, using, in Z.ai's words, "more environments, more diverse tasks, and more compute spent training on them." Lambert simplifies the difference between the two leading Chinese labs this way: Z.ai is strong at post-training, while Kimi is "more of a pretraining masterpiece."

The obvious question is how a model this small keeps up with the leading public American models, and whether the results are real. Lambert admits he has asked the same thing: "Surely the models aren't as good as they look." It bothers him that American companies with a huge resource advantage cannot pull away. The essay is his attempt at an answer.

A long lineage

His first answer is that Z.ai has been doing this longer than almost anyone. Zhipu AI was founded in 2019. The first GLM came out of Tsinghua University's THUDM group in March 2021, followed by GLM-130B in 2022, three ChatGLM chat releases in 2023, GLM-4 in January 2024 and GLM-5 in February 2026. The full timeline is on the entity page. Looking back on GLM-5.2, Lambert says it lived up to its hype. Weeks after release, researchers he knew still used it for its speed, some running it on internal clusters faster than public providers, and for its simplicity when working on frontier systems. The source dates GLM-5.2's release to June 22, but Lambert's own glm-5-2-step-change-for-open-agents puts the rollout on June 13 and the weights on June 16. June 22 is the date of that post.

Distillation is not the main explanation

The usual explanation is distillation. Lambert has written about it in the-distillation-panic and how-much-does-distillation-matter-for-chinese-llms, and he still thinks it is not the main factor. He accepts part of the concern. A recent paper, covered on stealing-reasoning-traces-from-proprietary-llm-apis, showed simple ways to extract reasoning traces from frontier model APIs, and Chinese labs could certainly use them at scale. What puzzles him is that US labs have not patched the leak and are asking the government for policy help instead. His counterargument comes from Z.ai's own description of the training. A lab that scales RL cannot simply distill its environments, the infrastructure to run them at scale, or the algorithms that combine them. how-distillation-is-used-today covers where distillation does help in an RL-heavy pipeline.

If distillation doesn't explain it, the next suspicion is benchmaxxing, meaning training toward test sets until real use falls short of the scores. Lambert's answer is a list of big-picture factors. He says technical details matter too, but they are harder to tell apart between labs.

Six reasons Chinese labs keep pace

He considers release speed the largest factor. Z.ai probably takes days to go from finished model to release, while OpenAI and Anthropic take months. Lambert thinks both American labs very likely have much better internal models than Z.ai or Moonshot. But they spend months on pre-release testing, and Chinese labs use that time to keep improving their benchmark scores. SpaceXAI, he thinks, behaves more like the Chinese labs here. This has been acceptable for American labs so far because demand for their models is still enormous. It could stop being acceptable if model self-improvement loops start depending on user data, because faster release cycles would then compound. Lambert calls these the race dynamics much of the industry worries about.

On benchmarks, he concedes Z.ai probably cares somewhat more about public scores than OpenAI or Anthropic do. Rankings like the Artificial Analysis Intelligence Index directly affect its stock price, its fundraising and its morale, since being the underdog that matches American giants is a good story. Subtle benchmaxxing is also normal across the industry. Many companies buy data specifically for the benchmarks they are behind on.

He does not think Z.ai has overtrained GLM-5.3 into a broken model, at least not on purpose, and expects Z.ai to check. Every lab is dealing with the rough edges of scaling RL. Anthropic's Opus 5 and Sonnet 5 have mixed reputations despite excellent scores (benchmarking-opus-5-slopcodebench has one attempt to measure Opus 5 on code quality). His view is that some model weights are easier to use than others, but the numbers in release blogs are real.

GLM-5.3 is probably a narrower model than Claude Fable or GPT Sol. OpenAI and Anthropic serve huge customers with endless use cases, while a company earlier in adoption can focus on the most valuable ones and make post-training easier by caring about less. Z.ai's flagship GLM models also have no vision, which helps their scores. Lambert tempers this with a Bloomberg report that Z.ai reached $1B in annual recurring revenue, mostly from on-premises deployments.

A fifth factor was added after the email went out. China's RL data industry is growing, driven largely by American data companies selling to Chinese labs. Chinese labs may be buying the same RL environments American frontier labs use and shipping the resulting models sooner. Lambert says the size of this market is still very uncertain.

The last factor is that Z.ai is simply very good, and probably much more compute-efficient than OpenAI or Anthropic. It has close ties to Tsinghua University and access to a large pool of eager talent, which matters as much to it as talent does to any Western lab. Lambert concludes that the GLM strategy is "a perfectly good" one and says he plans to test the weights through American inference providers like Fireworks or Baseten.

A staged release for cyber capabilities

The final section is about security, and here Z.ai itself changed how it releases models. It called GLM-5.3 its most capable model yet for cybersecurity, with large gains in vulnerability discovery, exploit analysis and multistep security tasks, and acknowledged "clear dual-use risks". It announced a staged release. Selected security partners would evaluate the model first in controlled settings, then broader access and the API would follow, and full weights would be published "once the necessary safety evaluations and release preparations are complete." On its own platform it described a request classifier and chain-of-thought monitoring on top of model alignment. That condition does not match the fixed two-week weights date at the top of the same post, and the source does not say whether the weights came out on schedule.

Lambert is sceptical that this kind of safety matters much. If GLM-5.3 does not ship these capabilities as open weights, another model will. Models with these capabilities keep getting smaller, and so easier to modify and deploy without safeguards. Diffusion is set by the least careful lab. He credits Z.ai with some good choices, including pushing proactive vulnerability discovery, but says no single company can handle this, and calls for industrial-scale guidance from government or industry coalitions to prepare all software for the change.

The staged programme resembles the one Thinking Machines describes in a-safe-path-to-open-weights: defenders get early access before the weights are released. Lambert's objection is the one openai-huggingface-incident-autonomous-hacking makes, that attackers already have comparable open models. It is also a long way from the Z.ai of 2025, whose product lead said in zai-playbook that the company open-sources a model "within a few hours" of finishing training.

Related pages

intelligence-vs-cost-linear prices GLM-5.3 at OpenRouter rates. It scores 60 on the Artificial Analysis index at $0.49 per task, against Fable 5.1's 66 at $3.69. qwen3-8-flash-next records the smaller GLM-5.3-Flash (320B total, 18B active) released alongside a Qwen preview. kimi-k3-open-weights-escalation is Lambert's earlier essay on capital efficiency, and this one extends it from pretraining scale to post-training scale.