Cactus Compute · The bottom of the inference market — a 14MB model and a dependency-free C++ runtime aimed at the four in five edge devices that cost under $200, where the unit economics are flash reads and milliwatt-hours rather than tokens.

They make AI models run on the device itself — phones, wearables, Raspberry Pis, microcontrollers — and hand a query off to a cloud model only when the small model knows it is out of its depth. Three products: Cactus Engine (a source-available C++ runtime and quantizer), Cactus Hybrid (a router post-trained *into* the model weights that scores its own confidence and escalates), and Needle 2 (their own 45M-parameter, 14MB tool-calling model that fits in 28MB of RAM).

Sector

Local & on-device AI

On the roster since

15 Sept 2026