AxionML / Mistral-Medium-3.5-128B-NVFP4

AxionML Mistral-Medium-3.5-128B-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 NVIDIA. The weights in this repository are an unmodified copy of nvidia/Mistral-Medium-3.5-128B-NVFP4 (revision b8c66d2098edd8c9c26bde2b2ff41b5967e111ae ). All credit for the quantization belongs to NVIDIA.

This is an NVFP4-quantized version of mistralai/Mistral-Medium-3.5-128B (128B parameters, dense), 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.

Use is governed by Mistral AI's Modified MIT License (included as LICENSE ): companies with more than US$20M global monthly revenue may not use this model without a commercial license from Mistral AI. NVIDIA's Open Model License also applies to the quantization (included as NVIDIA_OPEN_MODEL_LICENSE.md , with NOTICE ).

Model Summary

Evaluation Results

Scores reported by NVIDIA for this checkpoint. reasoning_effort="high" , temperature=0.7 , top_p=0.95 .

Quantization Details

• Quantization format: mixed — MLP linear layers in decoder layers 4–86 in NVFP4; MLP in layers 0–3 and 87, all attention linear layers in FP8
• KV cache: FP8
• Calibration dataset: blend from Nemotron-Post-Training-v3
• Tool: NVIDIA Model Optimizer v0.44.0

Usage

Deploy with SGLang

python3 -m sglang.launch_server \
    --model-path AxionML/Mistral-Medium-3.5-128B-NVFP4 \
    --tp 4 \
    --quantization modelopt_mixed \
    --trust-remote-code

Deploy with vLLM

vllm serve AxionML/Mistral-Medium-3.5-128B-NVFP4 \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.9 \
    --max-model-len 196608 \
    --config-format hf \
    --dtype auto \
    --trust-remote-code

Upstream validation was done with vllm/vllm-openai:v0.21.0 on B200; the SGLang command has not been validated upstream. Recommended sampling: temperature=0.7 , top_p=0.95 .

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: mistralai/Mistral-Medium-3.5-128B
• Quantization: nvidia/Mistral-Medium-3.5-128B-NVFP4 by NVIDIA
• Mirror: AxionML