From the source
Together AI announces ATLAS, a new adaptive-learning speculator system that improves inference performance at runtime, achieving up to 500 TPS on
DeepSeek-V3.1 and up to 460 TPS on Kimi-K2.
AdapTive-LeArning Speculator System (ATLAS): A New Paradigm in LLM Inference via Runtime-Learning Accelerators
Model update
From the source
ATLAS offers a new way of doing speculative decoding — one that dynamically improves at runtime — and it fits seamlessly alongside our other Turbo techniques like the proprietary Together Turbo Speculator or Custom Speculators.
together.ai
