Z.ai releases the weights of GLM-5.3, its strongest open model for code

Z.ai has published the weights of GLM-5.3 on Hugging Face under an open-weight license, making available for download a model that the company itself describes as the most capable in its line for code and long-horizon tasks. The repository had already racked up more than a thousand likes at launch and comes accompanied by the technical paper "GLM-5: from Vibe Coding to Agentic Engineering".

The central point of the update, according to the model card, is that GLM-5.3 uses the same base model as GLM-5.2: all the gain comes from post-training. In other words, this is not a new architecture or a new pretraining from scratch, but rather refinement on top of the previous base, focused on complex code and agent tasks that span many steps.

What changed compared to GLM-5.2

Z.ai claims a 50% improvement over GLM-5.2 on the internal Z.ai Code Bench benchmark and claims state of the art among open-weight models on public benchmarks such as Terminal Bench 3.0 and Agents' Last Exam. The numbers published in the card show significant jumps across several software engineering tasks:

| Benchmark | GLM-5.3 | GLM-5.2 | |---|---|---| | Terminal Bench 3.0 | 28.3 | 4.6 | | DeepSWE (v1.1) | 66.9 | 46.2 | | SWE-Marathon (v1.1) | 42.5 | 19.4 | | Terminal Bench 2.1 | 88.2 | 81.0 | | HLE with tools | 62.5 | 54.7 |

There is also a point the company raises with caution: an emerging cyber capability. According to the card, as post-training was scaled up, offensive security ability grew faster than expected, with the model reaching state of the art on CyberGym for vulnerability discovery (84.5, versus 77.2 for GLM-5.2) and more than doubling its predecessor on exploitation benchmarks such as ExploitBench (54.4 versus 24.4). This is a data point that matters both to those who work in security and to those who need to assess usage risks.

It's worth noting that the internal benchmarks and the comparison with proprietary models come from the model's own card, without independent verification so far. The full table from the source also includes competitors such as Kimi K3, DeepSeek-V4, Opus 4.8, and GPT-5.6 Sol, against which GLM-5.3 is competitive, though not always ahead of the top closed models.

Running without depending on a proprietary API

The appeal for the Brazilian developer lies in the word open-weight: the weights are published and the model can be served locally or by third-party providers, without being locked into a single closed API. The card lists support for several inference frameworks:

• SGLang, with its own cookbook
• vLLM, with recipes
• Transformers, via standard documentation
• KTransformers, Unsloth, and TokenSpeed
• Ascend NPU platform, with vLLM-Ascend, xLLM, and SGLang

One relevant operational detail: GLM-5.3 allows controlling the reasoning budget through the reasoning_effort parameter, which accepts low, high, and max (the default is max). To reproduce the benchmarks, Z.ai recommends keeping max; for chat scenarios, it suggests explicitly passing clear_thinking=true. This gives room to adjust cost and latency depending on the task.

On the size side, the repository's safetensors add up to more than 753 billion parameters in FP8 quantization, putting the total download at around 755 GB. It's not something that runs on just any machine, but through inference providers, access becomes trivial. On Hugging Face, the model already appeared with providers such as Baseten and zai-org itself, with an announced output price of US$4.4 and throughput of around 79 tokens/s on the former.

What the community is saying

In the Hacker News thread, the practical reception showed up quickly. User scosman summed up his impression of use:

"I've been using it more and more. Feels like Opus 4.8, in the best possible way."

>

-- scosman, on Hacker News

Another recurring point was the feasibility of running it locally and the savings via third parties. Commenter revolvingthrow argues that GLM-5.3 is "probably the sweet spot open weights model if you want to go beyond deepseek flash or the new glm flash," adding that it's "a lot easier to run" than Kimi, even though "slightly behind Kimi in ability," and that he expects pricing and speed "noticeably better" from third-party providers. He even notes the possibility of running the model quantized to 4 bits on a Mac with 512 GB of unified memory, with the caveat that the math might not add up.

On provider availability, fra pointed to DeepInfra as the first third-party provider on OpenRouter. pal9000i, meanwhile, raised a question that many people building products will have: how feasible it is to use an open model like this as a base for specialization via fine-tuning or LoRA for a specific use case. These are reader reactions, not verified facts, but they set the tone of the interest.

What remains open

For those developing in Brazil, the practical takeaway is that a model with declared performance close to top proprietary models on code tasks is available to download and serve on your own or through multiple providers, reducing API lock-in. What still needs independent validation are the benchmark numbers, which today come from the manufacturer's own card, and the exact policy of the glm-5.3 license, which is listed as "other" and deserves careful reading before any commercial use. The emerging cyber capability, finally, is a data point that requires extra attention from anyone putting the model into production.