GLM-5.3

ModelScope lists ZhipuAI/GLM-5.3 as a 753.33B-parameter Safetensors/PyTorch MoE model with sparse (DSA) attention, updated 2026-09-04 and totaling 755.66GB of weights. The model card duplicates the canonical Z.ai benchmark table and confirms open-source SOTA coding and CyberGym leadership. The page lists supported deployment frameworks (SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, plus Ascend NPU stacks) and the reasoning_effort parameter with low, high, and max values defaulting to max. It also notes 18,124 downloads on ModelScope, signaling parallel distribution alongside Hugging Face.

NVIDIA NIM's model card describes GLM-5.3 as a 753B-parameter text Mixture-of-Experts model from Z.ai, using DeepSeek-style sparse attention (DSA) over a 1,048,576-token context, with reasoning, function/tool calling, and a multi-token-prediction draft layer for speculative decoding. Release on Hugging Face is listed as 08/27/2026. Architecture details include GlmMoeDsaForCausalLM with 78 decoder layers (3 dense MLP plus 75 MoE) and 256 routed experts, corroborating the open-weights SOTA coding/cyber claims. Licensing notes an MIT-like GLM-5.3 License with a Z.ai security-review requirement for MaaS operators exceeding 10B USD over any 12-month period.

Z.ai's developer documentation for GLM-5.3 lists it as Z.ai's latest flagship, text-only, with a 1M-token context window and 128K maximum output. Reasoning is always on and supports low, high, and max effort levels, with disabling reasoning no longer supported and applications currently using thinking.type: disabled needing migration. The guide reiterates the 50% Code Bench gain over GLM-5.2 and open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam CLI, plus the CyberGym vulnerability-discovery lead. GLM-5.3 is available to all GLM Coding Plan users with the new pricing tier tied to the 50% coding improvement claim.

The Hugging Face release for zai-org/GLM-5.3 confirms the model is distributed open-weights and repeats the 50% Code Bench improvement plus open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam. It shares the full benchmark table against GLM-5.2, Kimi K3, DeepSeek-V4 Pro-0813, Qwen3.8-Max, Opus 4.8, Fable 5 (w/ fallback), and GPT-5.6 Sol across coding, cyber, and agentic suites. The card documents deployment support across SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, and Ascend NPUs via vLLM-Ascend, xLLM, and SGLang. It also notes a reasoning_effort parameter accepting low, high, and max levels, defaulting to max for controlling the thinking budget on long-horizon tasks.

Z.ai has released GLM-5.3, its latest flagship model, built on the same base as GLM-5.2 with all improvements coming from post-training scaling on the IndexShare, SAO, and slime stack. Z.ai reports a 50% gain over GLM-5.2 on its in-house Code Bench, open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam, and state-of-the-art CyberGym results for vulnerability discovery. Cyber capability emerged faster than expected during post-training, with exploitation-benchmark gains more than doubling GLM-5.2, including 54.4 on ExploitBench and 105/130 on ExploitGym (2h/6h). The official blog also notes weights would release on Hugging Face roughly two weeks after launch, subject to safety evaluation and hardening.