Inception Labs has unveiled Mercury 3, a diffusion-based large language model that has witnessed an impressive surge in its parameter count—now over 40 times greater than that of its original iteration. This leap has substantially bolstered its overall capabilities. In a series of benchmark tests, Mercury 3 has surpassed both GPT-6 Luna and Claude Haiku 5.5. Remarkably, its generation speed remains consistently high, producing thousands of Tokens per second, and at its peak, it operates approximately 8 times faster than GPT-6 Luna. Additionally, its context window has been broadened to accommodate 1 million Tokens.

Notably, Mercury 3 is priced significantly lower than its mainstream counterparts, offering a cost-effective solution. The model leverages parallel generation technology, demonstrating that the diffusion methodology can sustain its speed advantages even as the model scales up. This makes Mercury 3 highly versatile, adept at handling a wide array of tasks including knowledge-based Q&A sessions, interpreting complex instructions, and aiding in software engineering.

Currently, Mercury 3 is being utilized in various scenarios, such as powering browser agents and enabling real-time search functionalities for smart glasses. In Agent systems that necessitate multi-step operations and frequent calls, Mercury 3 stands out as an economical yet swift execution model. It efficiently manages high-frequency tasks like retrieval and information extraction, thereby markedly enhancing the overall efficiency of task completion while also reducing operational costs.