開放模型／AI 程式設計
GLM-5.3 Releases 744B MoE Weights, but Million-Token Deployment Still Requires Data-Center-Class Resources
Two weeks after launching the API, Z.ai has released the GLM-5.3 weights. The model retains GLM-5.2’s base architecture, with capability gains driven primarily by post-training for agentic coding and cybersecurity. The FP8 files still total about 756GB, and the custom license and vendor-led evaluations mean that performance, safety, and legal validation remain necessary before adoption.
Z.ai fulfilled its earlier promise on August 28 by uploading the GLM-5.3 weights to Hugging Face. This is not the previously reported GLM-5.3-Flash, which has 320B total parameters and activates 18B per token, but the full data-center-class version. The [model files](https://huggingface.co/zai-org/GLM-5.3/tree/main) consist of 141 safetensors shards, with the FP8 version totaling about 756GB. External reports list the model at 744B total parameters, with approximately 40B activated per token.
The configuration shows that the model has 78 layers and 256 routed experts, selecting eight experts per token in addition to one shared expert. Its native context limit is 1,048,576 tokens. It uses the GLM-5.2 base model, with its new capabilities coming primarily from post-training rather than another round of pretraining. According to the [official technical article](https://z.ai/blog/glm-5.3), the team used slime to connect Megatron for training, SGLang for rollouts, sandboxes, and verifiers, while continually adding long-horizon coding and cybersecurity environments. The model also offers three reasoning-effort levels—`low`, `high`, and `max`—and its deployment documentation covers vLLM, SGLang, Transformers, KTransformers, and Ascend NPU paths.
In Z.ai’s evaluations, the Terminal-Bench 3.0 score increased from 4.6 for GLM-5.2 to 28.3 for GLM-5.3, while DeepSWE v1.1 rose from 46.2 to 66.9. Its ExploitBench score increased from 24.4 to 54.4. However, the same table shows that it still trails closed models on several tests. Moreover, some scores, the agent harnesses, and the proprietary Z.ai Code Bench are all controlled by the development team, so they should not be treated as independently reproduced results.
Engineering teams must also account for two barriers. First, even when quantized, weights of this size do not make GLM-5.3 a typical workstation model. Full-throughput testing requires multiple accelerators, expert parallelism, and careful planning for the long-context KV cache. Second, GLM-5.3 uses a custom license. A [license summary and release report](https://gigazine.net/gsc_news/en/20260829-glm-5-3-open) states that users with annual revenue exceeding US$10 billion are subject to additional security-review conditions. The next questions are whether third parties can reproduce the coding and cybersecurity results using fixed versions of the serving stack, and what latency, VRAM consumption, and failure modes emerge under million-token workloads.