FP4

Xiaomi's MiMo-V2.5-Pro-UltraSpeed, a 1.02-trillion-parameter MoE model, achieved 1,000 tokens per second inference on standard cloud GPUs using FP4 quantization, DFlash speculative decoding, and TileR…

A study evaluating FP4 quantization-aware training (QAT) for real-time anomaly segmentation in brain tumor detection found that architecture choice, not QAT recipe, most determines model quality. The …

Researchers have developed ThriftAttention, a mixed-precision attention algorithm that selectively computes only 5% of query-key blocks in FP16 precision while processing the remaining 95% in FP4, rec…