Cactus

Tracked

An energy-efficient inference engine and numerical computing framework for phones, optimized for ARM CPUs to run large models with low power and memory footprint.

Author cactus-compute Open Sourced 2025-04-23 Last Commit Unknown

Cactus is a low-latency AI inference engine and numerical computing framework designed for mobile devices and wearables. Developed by cactus-compute, it is optimized for ARM CPUs to run large language models with minimal power consumption and memory footprint, enabling on-device AI without relying on cloud connectivity.

CPU-First Optimization

  • Tuned for ARM processors to reduce battery drain and heat generation during inference
  • Unified Cactus Graph and Cactus Kernels providing zero-copy computation graphs
  • SIMD-optimized kernels for high throughput on mobile hardware
  • Demonstrates higher CPU-only throughput and smaller model footprints compared to Llama.cpp on certain workloads

Cross-Platform SDKs

  • Flutter, React Native, and Kotlin SDKs for straightforward integration into any mobile application
  • On-device inference for chatbots, assistants, and quick generation tasks without network latency
  • Efficient deep learning inference embedded into mobile apps for real-time, privacy-preserving AI experiences
  • Hugging Face model conversion and benchmarking utilities to validate performance before shipping

API and Tooling

  • OpenAI-compatible C API with FFI bindings for integration across multiple programming languages
  • Python utilities for model conversion, testing scripts, and comprehensive build instructions
  • Rapid onboarding with complete documentation for mobile deployment workflows