AxionML / GLM-5.3-NVFP4

AxionML GLM-5.3-NVFP4

Mirrored by AxionML for open-source serving and deployment use cases. Part of AxionML's effort to provide ready-to-serve quantized models for the community.

Quantized by RadixArk. The weights in this repository are an unmodified copy of RadixArk/GLM-5.3-NVFP4 (revision 6e389189d564d9d3e4ea7fa284fe91f136d8ae2a ). All credit for the quantization belongs to RadixArk.

This is an NVFP4-quantized version of zai-org/GLM-5.3 (753B total parameters, ~40B activated), quantized with NVIDIA Model Optimizer .

About NVFP4 quantization: NVFP4 on Blackwell couples a compact E2M1 FP4 codebook with blockwise FP8 (E4M3) scaling over 16-element micro-blocks, so that 4-bit stored values remain numerically useful for neural-network computation. The E2M1 codebook provides a small, nonuniform set of representable magnitudes up to ±6 and relies on saturating behavior rather than IEEE NaN/Inf encodings to maximize usable range per bit. Using an FP8 block scale (rather than power-of-two-only E8M0) enables fractional scales and error-minimizing scale selection. On Blackwell Tensor Cores, native FP4 multipliers exploit E2M1 simplicity while higher-precision FP32 accumulation protects dot-product accuracy.

Ready for commercial and non-commercial use under the Z.AI Model License (MIT-style, included as LICENSE ).

Model Summary

Evaluation Results

Scores reported by RadixArk for this checkpoint. Against the BF16 source under the same protocol, GSM8K matched exactly and AIME 2026 pass@1 was within run-to-run noise.

Quantization Details

• Quantization format: NVFP4 W4A4 (group size 16, FP8 E4M3 block scales, static per-tensor activation scales) on the routed experts of all 75 MoE layers — 96.2% of parameters
• Unchanged (BF16): sparse attention incl. IndexShare indexer, shared experts, routers, the 3 dense MLP layers, norms, embeddings, lm_head , all MTP tensors
• Calibration dataset: 1,024 samples at length 512 from cnn_dailymail + Nemotron-Post-Training-Dataset-v2 , max calibration
• Tool: NVIDIA Model Optimizer v0.47.0.dev91

Usage

Deploy with SGLang

sglang serve \
    --model-path AxionML/GLM-5.3-NVFP4 \
    --tp-size 8 \
    --quantization modelopt_fp4 \
    --reasoning-parser glm45 \
    --tool-call-parser glm47 \
    --speculative-algorithm EAGLE \
    --speculative-num-steps 5 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 6

Validated upstream on 8x B300. The MTP layer is kept in BF16, so EAGLE speculative decoding works. See the SGLang GLM-5.3 cookbook .

Limitations

The base model was trained on data that may contain toxic language and societal biases. The quantized model inherits these limitations. It may generate inaccurate, biased, or offensive content. Please refer to the original model card and the upstream quantized model card for full details.

Credits

• Base model: zai-org/GLM-5.3
• Quantization: RadixArk/GLM-5.3-NVFP4 by RadixArk
• Mirror: AxionML