Together AI Ships ATLAS: Adaptive Speculative Decoding That Learns at Runtime, Claims 400% Inference Speedup

Together AI Blog·high signal

Together AI introduced ATLAS (AdapTive-LeArning Speculator System), the first speculative decoding system that dynamically adapts to workload patterns at runtime without manual tuning. Unlike static speculators that degrade on unseen inputs, ATLAS continuously learns — during a vibe-coding session it specializes for the specific code files being edited, increasing acceptance rates. Available now on dedicated endpoints at no additional cost to 800K+ developers.