In short
- The GLM-5.2 trails the Claude Opus 4.8 by only 1% on FrontierSWE—a benchmark for independent multi-hour projects—beating the GPT-5.5 on the same test. It is licensed under the MIT license with zero-region restrictions.
- The model was built entirely on Huawei Ascend chips with no NVIDIA hardware involved.
- Unsloth AI has already released 2-bit GGUF quantizations that reduce the sample size from 1.51TB to 238GB. You’ll still need 256GB of RAM or VRAM—but there, you can run it.
Z.ai has dropped GLM-5.2 on June 16, promising great performances, beating its previous GLM 5.1 lead.
A lab based in Beijing, which has been on the US Entity List from January 2025he seems to be benefiting from the growing concern over America’s approach to AI. Over the past week, the banning of Anthropic Fable and the release of this new model has helped drive zAI stock 90%, sending it to a new all-time high.
GLM 5.2 has numbers to support the hype.
On FrontierSWE—a benchmark that measures whether an AI assistant can complete measured technical tasks in hours, covering machine optimization, large code generation, and the use of ML research, the GLM-5.2 scored 74.4 against Claude Opus 4.8’s 75.1. It scored GPT-5.5 at 72.6. On SWE-bench Pro, which tests the independent resolution of GitHub’s world-class certification, GLM-5.2 scored 62.1 to GPT-5.5’s 58.6—and edged out GLM-5.1’s 58.4 by a margin.
The jump in quality makes it the best opening model to date in the Artificial Analysis Intelligence Index, which combines the results of 9 different factors to assess the quality of AI. OpenRouter software signs put it in the same category as the banned Claude Fable 5.
The tools used to achieve this are another interesting part of the story. GLM-5.2 was trained on Huawei Ascend chips—no Nvidia anywhere in the pipeline. Emad Mostaque, founder of Stability AI, comparison The total cost of education is about $25 million, 80% of it after education, which would make it much cheaper compared to its peers.
Like Decrypt reported earlier this yearZ.ai was already training image samples on Huawei’s Ascend Atlas servers without a single American chip. GLM-5.2 advances the development-the model of 744-billion-parameter-of-expert window of 1 million real images, five times the limit of 200K on GLM-5.1, and the permission of MIT which means that no government law can reverse the change of input.
Symbols and components of a tet model can read and create while Parameters are a set of internal settings and values that determine how the model processes information and creates responses.
Who it is and what it costs
For developers, the window of information is the change of operation. Repo-repo flows, multiple refactors, and long pipelines that need to be scaled are all a single call away. API pricing runs at $1.40 per million input tokens and $4.40 per million output – versus Claude Opus 4.8’s $5 input and $25 output. Coding plans start at around $18 per month and work directly within Claude Code, Cline, Kilo Code, and popular agent sites.
Local delivery is also technically possible. Free AI pushed 2-bit GGUF quantizations that compress the sample from 1.51TB to 238GB while maintaining ~82% accuracy.
Don’t get too excited, though. This means it needs 256GB of compatible memory or a similar RAM/VRAM-maxed M4 Ultra Mac Studio combo or workstation with a mid-range GPU and 256GB of RAM mixed with professional downloads. It’s still a lot of money, but something you can buy and drive from your home if you really want to.
We did a quick test, asking the GLM-5.2 to create our usual integrated shooter. The UI wasn’t very beautiful – some models made a nice appearance, but the experience was different: different events on the waves, the types of enemies that moved, bosses appearing later.
It created more different game worlds than anything else we tried for the same task at zero setup.
If you want to play it, it’s available here History of Itch.io.
This difference indicates where GLM-5.2 generates the most money. For many types of natural piping systems where variation is more important than polishing, the math on open price levels it’s hard to argue with that. At the most intense event – SWE-Marathon, where it scored 13.0 against Opus 4.8’s 26.0 – the difference in the closed margin is still real, and 13 wide.
Open-source weights are available HuggingFace under the MIT license. Quantized weights are also available at HuggingFace. Subscribers to the GLM Coding Plan can now switch to the GLM-5.2 sample cable, and it’s also available for free trial at. zI and other user issues.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
