Introducing Velvet-1: AI Model for Onchain Intelligence Today, we are commencing training for Velvet-1, our first flagship LLM for onchain intelligence. Onchain analytics is lagging behind coding and general knowledge reasoning capabilities. For truly effective onchain analytics, LLMs need to be constantly online: monitoring, detecting, and understanding data that is deeply interconnected. This must happen in real time, amidst high-velocity change. General-purpose models can reason about crypto, markets, and protocols in isolation, but they struggle to understand the full context of onchain & social systems. Their context windows and memory were designed and trained to work with static code and text data, not with onchain and social dynamics. We have developed new types of layers, context, and training techniques to handle these problems. The model is being developed on large-scale, continuously indexed datasets spanning: - Full onchain activity across major networks, including transactions, smart contract events, state changes, and mempool signals - Hyperliquid-specific market data, including perpetual orderbooks, funding rates, liquidations, trader positioning, and flow dynamics - Social and behavioral signals, including onchain and offchain graphs, wallet clustering, and profile embeddings Over the coming months, we will share technical updates, evaluation benchmarks, infrastructure notes, and early capability results as Velvet-1 progresses. Velvet-1 is our first step toward by-design on-chain intelligence.