~753B total / 40B active (MoE; same base as GLM-5.2, with the official GLM family specification also described as 744B-A40B)
1,048,576 tokens
GLM-5.3 License
Official open-weight release of Z.AI's GLM-5.3 flagship, published as a mixed-precision FP8 checkpoint across 141 safetensors shards totaling approximately 756 GB. GLM-5.3 uses the same base model and sparse Mixture-of-Experts architecture as GLM-5.2, with its reported gains coming entirely from expanded post-training on coding, long-horizon engineering, research, tool use, and cybersecurity environments. The official configuration specifies 78 transformer layers, a 6,144-dimensional hidden state, 64 attention heads, 256 routed experts plus one shared expert, eight routed experts selected per token, three initial dense feed-forward layers followed by sparse MoE layers, DeepSeek Sparse Attention with a top-k of 2,048, one Multi-Token Prediction layer, and a native 1,048,576-token context window. The released checkpoint uses dynamic blockwise FP8 quantization with selected modules retained in BF16 or F32. GLM-5.3 is an always-thinking model with low, high, and max reasoning-effort settings, defaulting to max; disabling thinking is no longer supported. Z.AI reports improvements over GLM-5.2 from 4.6 to 28.3 on Terminal-Bench 3.0, 46.2 to 66.9 on DeepSWE v1.1, 19.4 to 42.5 on SWE-Marathon v1.1, 26.2 to 48.2 on AutomationBench, and 77.2 to 84.5 on CyberGym. Official local-deployment guidance covers SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, and Ascend-oriented inference frameworks. The custom GLM-5.3 License broadly permits use, modification, redistribution, derivatives, and commercial deployment, while requiring Z.AI security review for Model-as-a-Service operators whose affiliated annual revenue exceeds 10 billion US dollars.