ROME
Web3 / ai data
ROME (ROME is Obviously an Agentic ModEl) is an open-source agent-specialised large language model built on a Mixture-of-Experts architecture and developed within an Agentic Learning Ecosystem (ALE). Unlike general-purpose LLMs fine-tuned for chat, ROME was trained from the ground up on over one million execution trajectories — recordings of agents actually completing tasks in real environments — using a framework called ROCK, an environment execution engine that provides secure, isolated sandboxes for agent actions. This training approach optimises ROME specifically for 'agentic crafting': natively understanding how to select and use tools, manage its own context window across long task horizons, and operate within complex multi-step workflows without losing track of objectives. Its open-source release gives researchers and developers a specialised foundation model that bridges the gap between raw language reasoning and actual software manipulation, without the overhead of prompt-engineering a general-purpose model into agentic behaviour. Example: A Web3 security team fine-tunes ROME on their own smart-contract audit trajectories — recordings of previous audits including tool calls, findings, and remediation steps — producing a specialised audit agent that reasons about Solidity contracts the way an experienced auditor does, rather than the way a generic text model does. Why it matters for AI and data in Web3: General-purpose LLMs require extensive prompt engineering to behave reliably as agents. ROME's trajectory-trained architecture makes it natively effective for the kinds of long-horizon, tool-using tasks — contract auditing, on-chain monitoring, governance analysis — that Web3 operators need to automate.
Explore the full Web3 Glossary — 2,062+ expert-curated definitions. Need guidance? Talk to our consultants.