Today, training advanced AI models is an exclusive club. It costs millions of dollars and demands specialized, tightly-controlled data centers with expensive, high-speed networking. This infrastructure barrier has created a "concentration problem," limiting who gets to innovate and who benefits.
Dispersed breaks through that barrier. We are mobilizing the world's most overlooked asset—the millions of powerful GPUs sitting idle in homes and offices—to create a global, open-source AI training network. If you have a GPU, you can contribute to training cutting-edge models and get paid for it.The Breakthrough: Training AI Over Standard Internet
The biggest obstacle to decentralized training is bandwidth. Traditional distributed training requires sharing tens of gigabytes of mathematical updates (gradients) every few seconds. Your home internet simply can't handle that—until now.
- 50x Bandwidth Compression: Our secret is a powerful compression algorithm that shrinks gradient updates by roughly 50 times. This means a training round that once needed gigabytes of bandwidth now only requires a few megabytes. A standard home broadband connection is all you need.
- Internet-Friendly Design: Forget firewall headaches and complex network setups. Node operators never connect directly; all communication is routed securely through cloud storage. No special hardware like InfiniBand or NVLink is required—just standard internet access.
Performance Built for the Real World
Dispersed is engineered for the unreliable, heterogeneous hardware of a global network:
- Always Contributing, Never Waiting (Asynchronous Training): Our network uses an asynchronous protocol, eliminating synchronization barriers. Node operators train independently and share results whenever they are ready. This means a super-fast RTX 5090 is never held back by a slower, compact device, even if the speed difference is 5x.
- Fault Tolerant: If a node goes offline, crashes, or a cloud spot instance is reclaimed, the training simply continues. Your prior contributions are already integrated into the model, ensuring zero work is lost.
- Universal Hardware Support: The system is smart enough to absorb any GPU—from high-end desktops to compact edge devices—without disruption.
Model-Agnostic and Future-Proof
The infrastructure supports any model architecture that learns through gradient descent, allowing different model types to run simultaneously without interfering with one another:
- Language Models: Train the next generation of text generators, including LLaMA, Mistral, and other transformer-based architectures.
- Image Generation Models: Contribute to diffusion transformers used for images (like Stable Diffusion and DALL-E families).
- 3D and Vision Models: Supports vision transformers, encoder-decoder, and convolutional layer architectures.
Zero-Trust Security & Fair Compensation
In an open network, trust is earned through transparency and cryptographic verification:
- Zero-Trust Security: Nodes operate under a strict zero-trust model. They never receive permanent storage credentials; the coordinator issues only short-lived, scoped access tokens.
- Integrity Checked: All contributions are cryptographically signed and integrity-checked before being applied to the model. A gradient validation step actively filters out corrupted or malicious data, protecting the model from bad actors.
- Solana Rewards: Every contribution is tracked as an "attestation" on the Solana blockchain. Rewards are automatically paid out based on the work done, weighted by your reputation score, which you build through consistent participation.
Get Started in a Single Command
The decentralized future of AI training is already here and running in production. Getting started is designed to be effortless:
- One Command: Run a single Docker command, and the container automatically detects your GPU, generates your cryptographic identity, and begins working. There is nothing else to configure.
