GLM-5.3-FP8-NVFP4

GLM-5.3 with NVFP4 routed experts and Z.ai's original FP8 weights for everything NVIDIA left in BF16.

How it was made

• Base: nvidia/GLM-5.3-NVFP4 @ e3b8f9f2fad29727dfa95e766c5cac18ea1414bf. Every tensor not listed below is copied byte-identical, including the NVFP4 experts (packed U8 weights, F8_E4M3 group-16 scales, F32 global scales and input scales) and NVIDIA's per-tensor FP8 dense MLP of layers 0-2.
• Replaced: 1435 tensors that NVIDIA stores as BF16 and zai-org/GLM-5.3 @ ba32f8c19c8738e2c57c71798c601d477a7bf844 stores as FP8 e4m3 with 128x128 block scales. Each is replaced by Z.ai's F8_E4M3 weight (same name and shape) plus its <name>_scale_inv F32 tensor of shape ceil(rows/128) x ceil(cols/128), copied byte for byte.
  • self_attn q_a_proj, q_b_proj, kv_a_proj_with_mqa, kv_b_proj, o_proj (79 layers), indexer wq_b and wk (22 layers)
  • shared_experts gate/up/down (layers 3-78)
  • MTP layer 78 routed experts, 256 x gate/up/down (NVIDIA left them BF16)
• Kept BF16 (BF16 in both): embeddings, lm_head, norms, router gate.weight, eh_proj, indexer weights_proj and k_norm. e_score_correction_bias stays F32.
• Tensor files keep NVIDIA's names (model-000NN-of-00050.safetensors); model.safetensors.index.json is regenerated. Tokenizer, chat template and generation config are NVIDIA's, unchanged.

Formats

• Routed experts, layers 3-77: NVFP4 as declared in quantization_config (modelopt, group size 16).
• Dense MLP, layers 0-2: FP8 per-tensor with scalar weight_scale/input_scale (NVIDIA's).
• Attention projections, shared experts, MTP experts: FP8 e4m3 128x128 block weights with weight_scale_inv (Z.ai's). quantization_config.fp8_block_quantization in config.json lists these modules.
• NVIDIA's hf_quant_config.json is omitted because it would not describe the FP8-block modules.

Size

438.4 GB total in 50 safetensors shards (438,379,016,352 bytes), versus 464.1 GB for nvidia/GLM-5.3-NVFP4.

License

Governed by the GLM-5.3 License (LICENSE, verbatim from Z.ai) and, for NVIDIA's quantization, the NVIDIA Open Model Agreement; see NOTICE.