Run compact AI on devices and route harder tasks to cloud inference when needed
Cactus Compute builds AI software for devices where memory, power and connectivity are limited. Its current product family includes an inference engine, hybrid routing and the compact Needle model. The offering is aimed at developers integrating AI into applications or hardware, not at home-service call answering under the separate Cactus brand.
Cactus Engine runs models on supported consumer and edge hardware, with quantization and hardware-aware execution. The Hybrid product adds a confidence-based route between local processing and a cloud model. A developer can keep simpler requests on the device while escalating harder cases according to the configured behavior. Cloud fallback means that a hybrid deployment is not automatically fully offline or fully local in its data handling.
Needle is designed around tool calling, device actions and structured extraction rather than open-ended general chat. The official Needle 2 page describes a compact model that maps a request to a defined function or schema. This narrower scope is useful when a device already exposes a limited set of actions and the main problem is choosing the right action and arguments from ordinary language.
The Hybrid router is not restricted to Cactus Engine; the vendor names other runtimes such as llama.cpp, MLX and Transformers. The engine page lists mobile and desktop support, while individual components target additional constrained devices. Developers should check the specific model, runtime and hardware combination rather than interpreting cross-platform marketing as identical performance on every device.
Cactus is advertised as free to start, but the engine is described as source-available under the Cactus license. That should not be casually relabeled as an unrestricted open-source license. A production evaluation should measure memory use, response quality, latency and fallback frequency using the application's real tool vocabulary. Teams also need to review model licenses, cloud-provider terms and which requests may leave the device before shipping the integration.
