Flash-MSA: Train on a million tokens without breaking the compute bank

A sparse attention kernel that makes long-context training finally affordable - if the benchmarks hold up outside the lab.

Jul 13, 2026 at 02:20 10 10