Deepgrove

DeepGrove builds edge-first AI foundation models that can be run on any device. Our models run on ternary weights (1) making them ~10x smaller, cut energy use by 80–90%,and run 2–20× faster on CPUs, NPUs, and GPUs — up to 100x faster on custom silicon.

Their current LLMs, Bonsai 0.5B, Lilac 1B, Aspen 4B match leading half-precision models of the same parameter count from the Qwen2.5, Llama 3.2, and Gemma 3 series. A preview checkpoint of our newest model Mangrove 30B, rivals models like Claude 3 Haiku and Llama 4 Scout at a fraction of the parameter count and can be locally deployed on an entry-level smartphone.

By redefining where AI can be used, Deepgrove is fueling a new market where cars, wearables, appliances, and industrial systems can run intelligence locally, enabling categories of products and services that were not previously feasible.

Sean met Eddie very early on when he was at Dartmouth, while hosting General Catalyst Venture Fellow's east coast presence at places like YC's 1st startup school in Boston, and throughout his time working on similar technology for Deepsilicon, a YCS24 startup.