GLM-5.3 model card (Z.ai)

Summary

Z.ai's model card for GLM-5.3, its newest open-weight model as of 2026-10-28. It is a 753B-parameter mixture-of-experts model on the same base as GLM-5.2, with all gains from post-training, aimed at long-horizon software engineering and agentic work. The weights come under Z.ai's own GLM-5.3 licence, which is MIT-style except for a security review for the very largest API providers. The card gives no release date; the model was announced on 2026-08-14 and a secondary report (GIGAZINE) puts the weights on 2026-08-29, which is the date used above.

Key points

• Size: 753B total parameters, model type glm_moe_dsa; the card does not give the active parameter count or a single context limit (its benchmark notes use 300K, 400K, and 1M tokens).
• Licence (GLM-5.3 License): free use, modification, redistribution, fine-tuning, and commercial use, keeping the copyright notice. One exception: a company that offers the model as a service (API inference or fine-tuning for third parties) and has more than US$10 billion in revenue with its affiliates over 12 months must pass a Z.ai security review before commercial use. Products that embed the model in features are not affected.
• Benchmarks (Z.ai's own figures): Terminal-Bench 2.1 88.2, Terminal-Bench 3.0 28.3, DeepSWE v1.1 66.9, CyberGym 84.5, Toolathlon Verified 73.0, HLE with tools 62.5, and 50% better than GLM-5.2 on its in-house Z.ai Code Bench.
• Running it: SGLang, vLLM, Transformers, KTransformers, Unsloth, and Ascend NPU back-ends. Thinking depth is set with reasoning_effort (low, high, or max; default max).
• Lineage: GLM-5 (2026-02, technical report arXiv 2602.15763), GLM-5.1 (2026-04), GLM-5.2 (2026-06, MIT licence per secondary reports), GLM-5.3 (2026-08).

Why it matters

GLM is one of the main Chinese open-weight families next to DeepSeek, Qwen, and Kimi. Its licence sits between MIT (DeepSeek) and Qwen's custom licence: open for almost everyone, with a review only for hyperscale API providers.