ruvnet/ruv-FANN

What it solves

ruv-FANN addresses the need for lightweight, CPU-native artificial intelligence that can be deployed without expensive GPU hardware. It solves the problem of resource-heavy AI by providing "ephemeral intelligence"—small, purpose-built neural networks that are instantiated for specific tasks and dissolved once completed to save memory and compute.

How it works

The framework consists of three integrated layers:

1. ruv-FANN Core: A high-performance Rust rewrite of the Fast Artificial Neural Network (FANN) library, providing the foundational neural network engine.
2. Neuro-Divergent: A forecasting layer featuring over 27 state-of-the-art models (including LSTMs and Transformers) optimized for speed and memory efficiency.
3. ruv-swarm: An orchestration layer that manages "swarms" of these lightweight networks, using WebAssembly (WASM) to run intelligence across browsers, servers, or embedded devices. It integrates with the Model Context Protocol (MCP) for use with tools like Claude Code.

Who it’s for

• Developers who want to integrate neural networks into applications without requiring CUDA or GPU infrastructure.
• Data Scientists looking for high-performance neural forecasting with lower memory overhead than Python-based alternatives.
• Edge Computing Engineers needing to run AI models in WASM environments or on RISC-V hardware.

Highlights

• High Performance: Claims an 84.8% solve rate on SWE-Bench and decision speeds under 100ms.
• Resource Efficient: CPU-native design that uses 29% less memory and 32.3% fewer tokens than baselines.
• Broad Compatibility: Supports 27+ neural architectures and 5 swarm topologies (mesh, ring, hierarchical, star, and custom).
• Memory Safe: Built entirely in Rust with zero unsafe code.