Open, Decentralized Artificial Intelligence
We believe that artificial intelligence should be a shared public resource, not a closed system controlled by a few.
Read more about why we are doing this.
This initiative exists to build a fully open-source AI ecosystem — from training algorithms to infrastructure to pre-trained models — developed transparently and owned by the community.
Importantly, our distributed network is designed to operate a single community-governed model and service. It is not a general-purpose public computing platform. Governance for training data, ethics, safeguards, and bias will be handled by a community process when the model reaches maturity.
Mission
Our mission is to create an AI stack where:
- Algorithms are open and auditable
- Pre-trained models are freely accessible
- Compute is distributed, voluntary, and decentralized — but narrowly scoped to serving and training a single open model/service
- Training is efficient without massive memory requirements
Every layer of the system is designed to be inspectable, reproducible, and community-owned.
Core Projects
🧠 Algorithms
Research into memory-efficient, decentralized training algorithms that work on heterogeneous machines and compose well with the distributed architecture.
🌐 Distributed Architecture
The Distributed Composable Neural Runtime (DCNR) is a permissionless, fault-tolerant compute network. It uses a distributed agent-based architecture with orchestrator-managed allocation for composable neural networks.
Explore the components:
- Orchestrator: Central coordination for network topology and node allocation.
- Physical Nodes (PNodes): Compute worker processes that host local execution.
- Virtual Nodes (VNodes): Stateful agents representing neural network components (layers, activations, cost functions).
- Gradient Locality: Distributed training without global coordination, using local parameters and gradients.
📦 Models
Fully open, reproducible pre-trained models built on top of the open infrastructure.
→ Models
Principles
- Open source is non-negotiable
- Transparency over convenience
- Decentralization over control
- Community ownership over profit
- Single-service scope over general-purpose compute
Roadmap & Status
Our development follows a compute-first roadmap: we start by building a decentralized distributed computing network, then progressively use it for training and inference.
The runtime is no longer just a design: multi-node distributed training works today, with GPU acceleration, mixed precision (fp16/bf16), a tokenized dataset service, and a live monitoring dashboard — all developed in the open in the compute-all repository.
Blog
Development notes and milestones, straight from the commit history.
→ Blog
Learn the Architecture
- Overview: Architecture Overview
- Components: Orchestrator · Physical Node (PNode) · Virtual Node (VNode) & Node Types · Protocol
- Building networks: Network Definition (YAML) · Production Training Networks
- Network layer: Network Membership & Discovery
- Security: Threat Model
- Reliability & trust: Trust & Validation
Join the Movement
This is an evolving project.
The design is not finished — and that is intentional.