mycellm: Decentralized GPU Sharing Network, Enabling Individual Computing Power to Participate in the AI Inference Ecosystem

mycellm integrates GPU resources distributed globally into a P2P inference network via the QUIC protocol and Ed25519 authentication. It allows users to both contribute computing power to earn credits and use open-source large models for free, without relying on cloud service providers.

The current large model inference market is monopolized by a few cloud service providers. Users either have to pay high API fees or own expensive GPU hardware themselves. Meanwhile, there are a large number of idle GPU resources worldwide—from personal gamers' graphics cards to laboratory training equipment—these computing powers are scattered everywhere and cannot form an effective supply network. The mycellm project was born to solve this contradiction: it integrates scattered GPU resources into a decentralized inference network through technical means, allowing direct connection between supply and demand sides of computing power.

mycellm's core vision is "Pool GPUs worldwide. Earn credits. No cloud required." The project uses a four-layer architecture design:

It provides multiple access methods, including iOS native apps, command-line chat interfaces, Web UI, and most importantly, OpenAI-compatible APIs. This means any tool that supports the OpenAI API format—such as Claude Code, aider, Continue.dev, etc.—can seamlessly switch to the mycellm network.

This is the core transport layer of the network. It uses the QUIC protocol instead of traditional TCP/HTTP to achieve lower latency and better NAT traversal capabilities. Combined with Kademlia DHT distributed hash table and STUN/ICE technologies, nodes can automatically discover each other and establish direct connections in complex network environments.

The underlying inference engine is based on llama.cpp, supporting multiple backends such as Metal (Apple Silicon), CUDA (NVIDIA), ROCm (AMD), and CPU. It also supports vLLM for higher throughput scenarios.

It uses an Ed25519-signed receipt system to generate encrypted proofs for each inference request, enabling a verifiable accounting mechanism, but it does not rely on blockchain or cryptocurrencies at all.

mycellm has designed a unique credit system:

• Seeder: Nodes running inference services earn credits by providing computing power
• Consumer: Users who use inference services consume credits to get model outputs
• Ed25519-signed receipt: Each request has an encrypted signature as an accounting voucher

The ingenuity of this design is that it not only incentivizes computing power contributions but also avoids the complexity and volatility of cryptocurrencies. Credits only circulate within the network and are completely isolated from fiat currencies or cryptocurrencies.