Modelpublic

modal-labs/glm-5.2-nvfp4-experts-only

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes152downloads

Model Card

GLM-5.2 NVFP4

Public NVFP4 quantization of zai-org/GLM-5.2.

Source checkpoint: zai-org/GLM-5.2 revision e32aaf0396e6987ee6dd2abb7f4d318b5f9b3cfe.

Quantization artifact: GLM-5.2-NVFP4-agentic-v2-b300-max-8k-512-r1.

Quantization scope: ONLY the routed MoE experts are NVFP4. Attention (MLA q/kv/o), shared experts, dense MLP (layers 0-2), DSA indexer, router gates, lm_head, and the MTP head all remain BF16