Flash: open-source model notes—smaller weights, sharper deploys
Distillation wins, KV-friendly stacks, edge stories worth tracking this round.
Offline viability and infra spend dominate mobile + edge teams, so disciplined small checkpoints and KV cache-friendly kernels stay fashionable.
The other thread pushes evaluation from leaderboard scores toward SLAs: under realistic jitter/concurrency are latency and defect rates predictable enough for commerce workflows?