Inception AI launches Mercury 3, claims speed and benchmark leads over rivals

Diffusion language model startup Inception AI has released Mercury 3, its new flagship model, claiming it outperformed OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5. 5 on most of its internal tests while generating responses at significantly higher speeds. The release addresses long-standing questions about the scalability of diffusion-based approaches in large language models.

According to benchmarks shared by the company, Mercury 3 achieved notably faster generation rates than its competitors in the same hardware environment, with Haiku 5. 5 and GPT-6 Luna trailing behind. The model’s context window has expanded to 1 million tokens, a substantial increase from the 260,000 tokens supported by its predecessor, Mercury 2. 5. While the exact parameter count remains undisclosed, Inception stated that Mercury 3 is over 40 times larger than the original Mercury 1.

The strongest performance advantage appeared in coding-related evaluations, where Mercury 3 scored higher than both GPT-6 Luna and Haiku 5. 5 on one benchmark. It fell slightly behind Haiku 5. 5 on another coding test and on a long-context evaluation, while GPT-6 Luna edged it out in two other categories by narrow margins. No third-party verification of these results has been published to date.

Unlike most large language models, which generate responses sequentially, Mercury 3 produces an initial draft of the entire response and refines it in parallel. Co-founder and CEO Stefano Ermon suggested that the primary concerns about diffusion models have centered on scalability, framing the constraints as resource-related rather than inherent to the architecture. The company has already begun training a successor to Mercury 3, indicating confidence in the approach’s long-term viability.

Early users are testing the model in applications where response time is critical. Kuan Zou, founder of real-time search startup Octen AI, noted that Mercury 3’s median generation time met the demands of latency-sensitive use cases like smart glasses, where delays can disrupt the user experience. At tutoring platform RevisionDojo, CTO Michael Tong reported similarly fast response times for hints and marked answers.

A temporary promotional rate further reduces costs, positioning the model competitively against other offerings in the market.

Developers can access Mercury 3 through Inception’s platform or via integrations with third-party providers. The model is designed to work with existing OpenAI API-compatible systems with minimal adjustments. Inception, founded in 2024 by Ermon and fellow computer science professors Aditya Grover and Volodymyr Kuleshov, raised $50 million in a funding round led by Menlo Ventures last November, with participation from Microsoft’s venture arm.

Earlier reports suggested acquisition interest from major tech companies, with Inception reportedly seeking a valuation above $1 billion. No deal has been confirmed, leaving the startup’s future trajectory uncertain.

The results, if validated externally, could prompt further exploration of diffusion models as an alternative to traditional architectures. Speed has become an increasingly important factor in applications like real-time search, augmented reality, and interactive tools, where even small delays can impact usability. Whether these advantages hold in broader deployment remains to be seen, as independent testing will be key to assessing the model’s real-world performance.

For now, the release adds another dimension to the ongoing competition in AI model development, highlighting the trade-offs between different approaches to generation and scalability.

Sources: siliconangle.com