Ultra-fast inference on custom LPU hardware. Runs open-weight models at hundreds of tokens per second.