NVIDIA has released a model quantization in NVFP4 format, based on the Mixture-of-Experts (MoE) language model "DeepSeek-V4-Pro-0813" developed by DeepSeek AI.

This model was quantization using NVIDIA's Model Optimizer. Notably, the draft heads of the DSpark speculative decoding module used to accelerate inference have also been losslessly converted from MXFP4 to NVFP4. The model is configured so that both the target model and the draft model maintain consistent quantization settings within a single checkpoint.

The model supports inference in SGLang and vLLM. NVIDIA explained that this model was not developed by NVIDIA, but was quantization and packaged by NVIDIA for specific use cases.

Source: