Lamb Labs

AI inference today runs on general-purpose GPUs that must move every model weight through a general-purpose stack, creating a memory-bandwidth bottleneck. This limits inference speed for large language models. Companies and developers wanting faster token generation are constrained by this architecture, which was not designed specifically for running a single fixed model as fast as possible.

Lamb Labs builds MPUs (Model Processing Units) that hardcode an entire model, including its weights, into silicon so the model itself becomes the chip. Their approach moves through four stages: a one-bit quantized version of Qwen3.8 27B called Larry (3.8GB, 14.6x smaller than the original, retaining 92% of Qwen's IFEval strict-prompt score) running live on a single GPU; an FPGA accelerator running an 8B-parameter model under 10W on a Kria KV260 board; Little Lamb, a self-trained tiny language model burned into FPGA fabric that generates tokens entirely on-chip with no DRAM in the loop (12,000 FPGA token steps/sec on a $250 board); and the endgame of custom ASICs targeting 20,000+ tokens per second for a 27B-parameter model, using an RL environment to search chip designs for speed and energy efficiency per token.