Sakana AI–Modal Namazu collaboration reportedly serves a roughly 1T model across four Blackwell GPUs

A post on the Sakana AI–Modal Namazu serving collaboration says a roughly 1T model runs across four Blackwell GPUs in Modal’s serverless pool. It reports that a custom draft model for speculative decoding, trained on 30B-plus token IDs from real traffic, brought 25% higher acceptance, roughly 35% lower per-token decode latency and about 40% more traffic per replica. It says post-launch DFlash2 retraining raised average acceptance length from 3.51 to 5.01 tokens.