hyperspaceai/hyperspace-node

What it solves

Hyperspace creates a decentralized, peer-to-peer (P2P) network for AI inference, removing the need for central servers. It allows users to contribute their own hardware resources—such as GPU compute, storage, and bandwidth—to a global pool that can be used to run AI models, while earning rewards for their contributions.

How it works

Built on the libp2p protocol stack, the network uses a Distributed Hash Table (DHT) for peer discovery and GossipSub for communication. Agents can be deployed as a CLI tool, a desktop tray app, or directly in a browser via WebGPU.

Depending on their hardware, agents provide different "capabilities," including:

• Inference: Running LLMs (supporting GGUF models up to 32B parameters).
• Embedding: Generating vector embeddings on CPUs.
• Storage & Memory: Providing distributed block storage and a vector database.
• Relay & Proxy: Helping other agents connect and providing residential IP proxies.

To ensure agents are honest about their resources, the network uses a "Pulse" verification protocol—a heartbeat system using Merkle proofs to verify that agents are actually online and capable of performing work.

Who it’s for

• Hardware owners who want to monetize their idle GPU or CPU power.
• AI developers and users seeking a decentralized alternative to centralized AI inference providers.
• Users with limited hardware who can leverage the P2P network to run larger models than their local machine supports.

Highlights

• Multi-platform support: Runs on macOS, Linux, Windows, and directly in the browser.
• Incentive system: Rewards users with points based on uptime (presence) and actual work performed (active tasks).
• Flexible deployment: Offers a CLI for headless servers, a tray app for desktop users, and a zero-install browser version.
• Native integration: Works with llama-server and Ollama for optimized GPU inference.