GLM-5.3-NVFP4

This repository contains an NVFP4-quantized version of Z.ai's zai-org/GLM-5.3. The weights and activations of the routed-expert linear layers are in NVFP4, and the KV cache is in FP8. The checkpoint uses the NVIDIA Model Optimizer format (quant_method: modelopt) and is served directly by SGLang and vLLM on NVIDIA Blackwell GPUs (SM100+).

• Base model: zai-org/GLM-5.3
• Quantization: NVFP4 weights + static NVFP4 activations; FP8 KV cache
• Size: 433 GiB
• License: GLM-5.3 License

Quantization Method

This checkpoint was produced with our in-house post-training quantization toolchain and exported in the NVIDIA Model Optimizer format. Only the weights and activations of the linear operators within the routed MoE experts are quantized to NVFP4, with static per-tensor activation scales. The KV cache is quantized to FP8 with static unit scales (kv_cache_quant_algo: FP8).

Quick Start

The commands below enable speculative decoding with the DFlash 2 draft model (lossless, 7 draft tokens per verification step); remove the --speculative-* flags to serve without it.

Serve with SGLang (main):

python3 -m sglang.launch_server \
    --model-path incoai/GLM-5.3-NVFP4 \
    --tp 8 \
    --quantization modelopt_fp4 \
    --reasoning-parser glm45 \
    --tool-call-parser glm47 \
    --chunked-prefill-size 8192 \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path incoai/GLM-5.3-DFlash2 \
    --speculative-draft-attention-backend trtllm_mha

Or with vLLM (v0.28.0 or later):

vllm serve incoai/GLM-5.3-NVFP4 \
    --tensor-parallel-size 8 \
    --reasoning-parser glm45 \
    --tool-call-parser glm47 \
    --enable-auto-tool-choice \
    --speculative-config '{"method":"dflash","model":"incoai/GLM-5.3-DFlash2","num_speculative_tokens":7}'

Accuracy

We compare Z.ai's original FP8 release with this NVFP4 checkpoint. Higher is better.

License

This model is a quantized version of GLM-5.3 and is distributed under Z.ai's GLM-5.3 License, which it inherits from the base model.