Speaking of AI projects in the Sui ecosystem, there have been quite a few emerging in the past two years, but how many can truly bring AI agents onto the chain to run in practice? Very few. Most projects remain at the proof-of-concept stage—large model parameters cause lag, inference delays explode, data privacy issues remain unresolved, and ultimately AI becomes just an ornament for DeFi, gaming, and social media, not the main focus.

Talus AI is different. Starting from late 2025, they directly treat Walrus as infrastructure, deploying large-scale models, dynamic datasets, and on-chain historical records, enabling agents to make real-time decisions, trade, and interact. The entire process is decentralized and fully verifiable—this is not minor tweaks but a reimagining of Web3’s intelligent interaction methods.

The core of Talus is AI agents, essentially digital avatars. Users define their goals, and the agents automatically handle the tasks—DeFi arbitrage, gaming gear farming, social posting and interaction, all possible. Previously, agents ran on centralized servers, operating as black boxes—no one knew what was happening. Now, it’s different: models reside in Walrus, decision records are on-chain, and the entire execution process is transparent and traceable. This transforms agents from untrusted to trustworthy, with every step auditable and users able to verify the execution path.

Walrus’s key role here is as a data brain. Agents need to process massive amounts of on-chain transaction history, market signals, and user preferences. These data are scattered everywhere and extremely large, making traditional storage impossible. Talus organizes data using Walrus’s blob objects, with model parameters stored hierarchically—small layers in Quilt, larger layers stored directly—when querying, Sui contracts first confirm blob availability, then fetch in parallel.