NVIDIA Releases NVFP4 Quantized MiniMax-M3 AI Model on Hugging Face
NVIDIA has released a quantized version of the MiniMax-M3 multimodal AI model on Hugging Face, making the 428-billion parameter mixture-of-experts model more accessible for AI developers and researchers. The release features NVFP4, a 4-bit floating-point compression format developed by NVIDIA to reduce the memory footprint of large language models while preserving inference accuracy.
Compressing MiniMax-M3 to 4-bit precision allows the model, which supports up to a 1-million token context window, to run with 2x the memory efficiency on NVIDIA's Blackwell GPU architecture. Multimodal MoE models generally require massive amounts of memory and compute to load and operate. By lowering the precision requirement, NVIDIA enables developers to deploy larger models more easily on their existing hardware without downloading the full uncompressed weights.
From the sources (2 posts)
@_akhaliqRT @HuggingPapers: NVIDIA just released an NVFP4-quantized MiniMax-M3 on Hugging Face A 428B parameter multimodal MoE model with a 1M-tok…
@huggingpapersNVIDIA just released an NVFP4-quantized MiniMax-M3 on Hugging Face A 428B parameter multimodal MoE model with a 1M-token context window, now compressed to 4-bit precision for 2x memory savings on Blackwell GPUs.