Z.ai's GLM-5.3 weights are a 753B fp8 MoE on a brand-new architecture.

The repo config declares GlmMoeDsaForCausalLM, model_type glm_moe_dsa, 8 experts per token, fp8. 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. 1,210 likes against only 8,804 downloads, which matches the 8-GPU self-hosting requirement. Most builders will meet this through an API. (Hugging Face)