Problem

AI applications that depend entirely on the cloud can be slow, expensive, privacy-sensitive, and unusable when connectivity is poor. Every request that travels to a server adds round-trip latency and inference cost, while sensitive audio, images, or text may leave the user’s device. This is especially painful for mobile, wearable, and edge applications that need responsive behavior, offline operation, or strict data controls.

Cactus’s clearest use case is real-time voice interaction on phones and other edge devices: voice assistants, dictation, and transcription need to respond in roughly real time, while noisy audio or complex requests may still require cloud-scale models. The company also targets private RAG, phone-control agents, and medical applications where local processing is valuable.

Product / Service

Cactus provides an open-source, cross-platform inference engine and SDK for running language, vision, speech, and embedding models on smartphones, laptops, wearables, and other low-power hardware. Developers can integrate it through bindings and APIs for platforms including iOS, Android, macOS, Flutter, React Native, Kotlin, Swift, C++, and Python. The engine supports model optimization techniques such as quantization and NPU acceleration, with support for a broad range of model families.

Its differentiator is a hybrid edge-cloud architecture. Cactus runs straightforward requests locally for low latency, privacy, lower cost, and offline resilience, then automatically hands more difficult or noisy requests to cloud models based on confidence or operating conditions. The company claims sub-120-millisecond on-device latency, more than 5x cost savings, and that over 80% of production transcription and LLM inference can be handled on-device; its delivery model combines a free-starting developer toolkit with optional cloud fallback rather than requiring every workload to run remotely.

Market

Cactus competes in on-device AI inference, edge AI, and developer tooling, with an emphasis on a unified, cross-platform runtime rather than a platform-specific framework. Its named alternatives include llama.cpp, ExecuTorch, MLC LLM, ONNX Runtime, TensorFlow Lite, Core ML, MediaPipe, MLX, whisper.cpp, Nexa AI, Argmax, and Liquid AI. Cactus positions itself around combining mobile deployment, multimodal models, open-source access, and automatic cloud fallback.

The company is an active Y Combinator Summer 2025 startup based in San Francisco with an eight-person team. Its public GitHub repository has more than 4,200 stars, and the company says customers have built private RAG pipelines, offline fallbacks, phone-tool agents, and privacy-sensitive AI applications. Public materials do not disclose revenue, funding totals, or named customers, so its commercial stage cannot be confirmed as pre-revenue; the available evidence supports describing it as an early-stage but publicly released and actively adopted developer platform.