Zhipu AI released GLM-5.3-BF16 open-weights model on Hugging Face
Z.ai (Zhipu) released GLM-5.3-BF16 on Hugging Face. GLM-5.3 uses the same base model as GLM-5.2, with gains from post-training. On Terminal Bench 3.0, GLM-5.3 scores 28.3 vs GLM-5.2's 4.6. On CyberGym, GLM-5.3 scores 84.5 vs GLM-5.2's 77.2. On ExploitBench, GLM-5.3 scores 54.4 vs GLM-5.2's 24.4. GLM-5.3 supports deployment via SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, and Ascend NPU frameworks (vLLM-Ascend, xLLM, SGLang).
China context
- Original name
- 智谱
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download and deploy GLM-5.3-BF16 from Hugging Face using supported frameworks like vLLM or SGLang.
- For investors
- The release of GLM-5.3 with competitive benchmark scores may intensify open-weights competition among Chinese AI labs.
Zhipu AI released GLM-5.3-BF16, an open-weights text-generation model, on Hugging Face. The model shares the same base as GLM-5.2 but improves post-training, with notable gains in coding and cyber capability benchmarks.
GLM-5.3 shows a 50% improvement over GLM-5.2 on Z.ai Code Bench and large gains on Terminal Bench 3.0 (28.3 vs 4.6) and ExploitBench (54.4 vs 24.4). The model supports a reasoning effort parameter with low, high, and max levels, defaulting to max.
The release positions GLM-5.3 as a strong open-weights competitor in coding and cyber tasks, with benchmark scores approaching or exceeding some proprietary models like GPT-5.6 Sol on Terminal Bench 3.0 (28.3 vs 34.6).
For builders outside China, GLM-5.3 offers an open-weights model with strong coding and cyber capabilities, deployable on multiple frameworks including Ascend NPU. For investors, the release signals Zhipu's continued investment in open-weights competitiveness.
Observable next signals include adoption of GLM-5.3 in developer tools and security workflows, and whether other Chinese labs respond with updated open-weights models.