We have developed a foundational Crypto World Model (Theia) on the Chainbase omnichain data network. Theia aims to transform vast amounts of on-chain dark knowledge into comprehensive intelligence for humans in a trustworthy manner. This is accomplished by distilling knowledge from both language models and the Chainbase omnichain data network through model training.

How do we build Theia?

We propose a model tuning algorithm for Theia, Decentralized Weight-Decomposed Low-Rank Adaptation (D2ORAD^2ORA) technique by revamping the DORADORA by NVIDIA, which decomposes a large language model into magnitude and direction components in a trustworthy manner. By feeding the model with on-chain data, we derive the crypto magnitude vector (McM_c) and the directional matrix (DcD_c). The product of these components represents a crypto-oriented parameter matrix that encapsulates the extensive knowledge within the crypto space. The decentralized training approach ensures data security, model robustness, and result transparency. Consequently, the crypto knowledge is condensed into the crypto-oriented matrix, independent of the large language model (LLM), allowing it to be enhanced with future data or on-chain activities, thus evolving into an all-in-one expert in crypto.

Capabilities of Theia

Unlike existing crypto AI projects, Theia offers a complete reasoning chain for question answering, providing transparent and trustworthy results. It learns crypto patterns from extensive on-chain and off-chain data and spatial-temporal activities, and its responses are based on causal reasoning with these crypto patterns. Ultimately, Theia consists of a powerful LLM parameter matrix operating with the crypto-oriented parameter matrix and crypto patterns. To ensure up-to-date intelligence with reliable arguments, we integrate Theia with the real-time omnichain data network and an Internet AI reader via Retrieval-Augmented Generation (RAG). Leveraging Theia’s intelligence, users can also build task-specific AI models on top of it.


Theia represents the next generation of Crypto World Models, derived from model training using the novel D2ORAD^2ORA algorithm and crypto patterns identified by machine intelligence. It offers transparent and trustworthy knowledge to the web3 community, illuminating the hidden knowledge within the crypto world.