OSS
Z.ai put GLM-5.3's weights on Hugging Face as a 753B fp8 MoE with a brand-new glm_moe_dsa architecture
zai-org/GLM-5.3 went live on Hugging Face on 2026-08-25 and hit the HN front page at 717 points on 2026-08-28; the repo config declares GlmMoeDsaForCausalLM, model_type glm_moe_dsa, 8 experts per token, and fp8 quantization, with 1,210 likes against only 8,804 downloads. Z.ai says the model reuses the GLM-5.2 base and gets all its gains from post-training, claiming 28.3 on Terminal-Bench 3.0 (up from 4.6), 88.2 on Terminal-Bench 2.1, and 84.5 on CyberGym. The low download count relative to likes matches the 8-GPU self-hosting requirement secondary coverage reports, so most builders will hit this through the API rather than locally.
Source
↳ Follow the thread