NVIDIA uploaded Qwen3.8-Flash-Next-NVFP4-QAD to Hugging Face, providing quantized image-text MoE weights with approximately 180B parameters. This checkpoint underwent 100 optimizer updates and combines NVFP4 and MXFP8, with the weight files totaling approximately 129.7 GB. The model card states that QAD involves quantization-aware training rather than simply requantizing NVFP4 weights, and warns that the accuracy gains have not yet been statistically supported.
Source:explainx.ai · Read original
Leave a Reply