Better LLM pre-training in NVFP4

At Data Fest 2026 in Belgrade, Andrei Panferov from the Institute of Science and Technology Austria introduced Quartet II, a novel method for NVFP4 pre-training that recovers SOTA accuracy. He outlined the core challenges of low-precision LLM training and presented CUDA kernels tuned for Blackwell GPUs, ready for integration into real training pipelines. #MultimodalAI