cactus-compute/cactus
Quantization, kernels, runtime and inference engine for mobiles, wearables, smart home and robots.
What it solves
Cactus is a hybrid edge-cloud AI engine designed specifically for mobile devices and wearables. It enables high-performance local inference for text, speech, and vision models while providing a mechanism to automatically hand off complex queries to the cloud when local model confidence is low.
How it works
The system is built as a layered stack:
- Cactus Engine: Provides OpenAI-compatible APIs for multimodal AI tasks and manages the hybrid cloud handoff logic.
- Cactus Graph: A zero-copy computation graph for efficient tensor operations.
- Cactus Kernels: Optimized CPU/GPU kernels (ARM NEON SIMD) tailored for mobile hardware like Apple, Samsung, and Pixel devices.
- Cactus Quants: A custom rotation-based quantization technique that supports weight tensors from 4-bit down to 1-bit.
- Cactus Transpiler: A Python-based tool that converts PyTorch models into the Cactus runtime graph format.
Who it’s for
Developers building AI-powered mobile and wearable applications who need low-latency, offline-capable inference with the option of cloud fallback for higher accuracy.
Highlights
- Hybrid Execution: Automatically routes hard queries to the cloud based on local model confidence thresholds.
- Multimodal Support: Local execution of LLMs, VLMs (Vision Language Models), and transcription models (e.g., Whisper, Parakeet).
- Broad Hardware Optimization: Specialized kernels for mobile CPUs and GPUs across various Android and iOS devices.
- Extreme Quantization: Supports mixed-precision and uniform quantization from 4-bit to 1-bit.
- Cross-Platform Bindings: Official bindings for Swift, Kotlin, Flutter, React Native, Python, and Rust.