Sentient unveils 'ROMA V2,' a high-efficiency AI agent engine…cutting inference costs and boosting performance

Sentient, a decentralized AI (DeAI) infrastructure project led by Sandeep Nailwal, co-founder of Polygon, said on the 10th that it has unveiled 'ROMA V2,' an AI agent engine that simultaneously improves cost efficiency and inference performance.

The company aims to offer an alternative to a market structure dominated by high-cost AI models, by adopting an architecture that lowers inference costs while improving performance versus existing systems. The release marks an achievement one year after Sentient raised seed funding of about $85 million (approximately 123 billion won).

At the core of ROMA V2 is a 'Self-correcting Reasoning' feature that enables the AI to check for and fix logical errors on its own before generating an answer. Sentient said the system is designed to proactively improve the accuracy of outputs. According to results from the company's reasoning test (SEAL-0), ROMA V2 posted roughly a 10% improvement in problem-solving performance compared with existing models.

In particular, Sentient's internal tests estimated that combining ROMA V2 with the DeepSeek-V3 model could deliver performance comparable to leading paid AI models such as Claude 3.5 Sonnet, while cutting inference costs by up to more than 80%. The company believes this architecture could significantly lower the barriers to adopting high-performance AI.

A structure aimed at improving transparency in the AI decision-making process was also introduced. ROMA V2 records the basis for the AI's judgments in data form and structures the reasoning path as a 'digital receipt,' allowing users to verify it on-chain later without tampering. The mechanism is designed to let users directly validate the rationale behind specific AI decisions, with a focus on mitigating the opacity of AI decision-making.

Sentient said the ROMA V2 launch will serve as a springboard to accelerate the expansion of the decentralized AI ecosystem. A Sentient official said, "ROMA V2 goes beyond a simple performance upgrade; it is designed with safety and reliability in real-world service environments in mind," adding, "Our goal is to establish a standard for safe AI agents that people can trust and use in everyday life, including a function that enables the AI to detect an error and halt before an incorrect payment is executed."