We are a small team of kernel engineers and systems researchers. We build the software that unlocks the performance already inside every chip.
Qwen3.8-27B on one RTX PRO 6000 Blackwell. Time to first token at 256K.
Than cuDNN, in 13 rounds for about $1. No manual tuning.
Run log on GitHub →Profiling, rewriting and validating, for weeks, on each chip.
The same loop, run autonomously across models and chips.
The latest open-weight models, across the accelerators you already run.