Introducing Velvet-1
Building an AI Model for Onchain Intelligence
Large language models have transformed how we write code, reason through complex problems, and interact with information. But one area has remained largely unsolved: understanding onchain systems in real time.
Today’s AI models were built for a different world. They excel at reasoning over static information such as code, documentation, books, and web pages. Blockchain data is different. It’s naturally represented as massive, interconnected graphs, where every wallet, transaction, smart contract, protocol, liquidity pool, and token is connected to countless others through relationships that are constantly shifting.
These graphs evolve every second. New transactions settle, liquidity moves, positions open and close, and entirely new networks of interaction emerge continuously. General-purpose LLMs were designed to learn from sequences of words and static documents, not dynamic graph structures that evolve in real time.
Understanding what’s happening onchain requires more than retrieval. It requires reasoning over evolving graph structures, and understanding the full context of onchain activity, market structure, and social dynamics as one interconnected, continuously evolving system.
Building a model capable of this kind of reasoning requires enormous volumes of onchain data, and sustained research into new architectural layers, context mechanisms, and training techniques designed specifically for continuously evolving onchain environments.
Training on the World’s Largest Onchain Context
Velvet-1 is being trained on continuously indexed datasets spanning multiple dimensions of the crypto ecosystem.
Full Onchain Activity
Across major blockchain networks, including:
Transactions
Smart contract events
State changes
Mempool signals
This gives the model visibility into how protocols and markets evolve in real time.
Hyperliquid Market Intelligence
Including:
Perpetual order books
Funding rates
Liquidations
Trader positioning
Order flow dynamics
This helps Velvet-1 develop a deep understanding of market structure and derivatives trading.
Social & Behavioral Intelligence
Markets aren’t driven by transactions alone. They’re driven by people.
Velvet-1 is also trained on rich behavioral datasets, including:
Onchain social graphs
Offchain social signals
Wallet clustering
Profile embeddings
This allows the model to understand how narratives, communities, and capital move together.
The Road Ahead
Training has only just begun. Over the coming months, we’ll share technical updates as development progresses, including:
Architecture insights
Training methodologies
Infrastructure updates
Evaluation benchmarks
Early capability demonstrations
This is only the beginning. Velvet-1 is our first step toward AI built by design for onchain intelligence, a new foundation for autonomous agents, smarter trading systems, and the future of SocialFAI.