Mercury 3: Diffusion at scale

The cost to run real-time LLM applications drops significantly, enabling more complex AI agents and interactive services.

Inception has launched Mercury 3, a new diffusion-based large language model (LLM) that demonstrates significant performance improvements over existing models. Mercury 3 achieves speeds of over 1,000 tokens per second, outperforming models like GPT-6 Luna and Claude Haiku 5.5 while running 4-8x faster. The model boasts a 1 million token context window, a substantial increase from Mercury 2.5's 260K tokens. Inception highlights that Mercury 3 scales effectively, being over 40 times larger than its predecessor, Mercury 1, yet maintaining its speed and vastly improving intelligence. Early access partners are already utilizing Mercury 3 for applications including voice and chat support, real-time tutoring, data engineering, and web navigation. The model is available via the Inception API platform with introductory pricing and compatibility with existing OpenAI API stacks.

Mercury 3's performance indicates that diffusion-based LLMs are now a viable and scalable alternative to autoregressive models for demanding, latency-sensitive tasks. Its speed and context window improvements directly address key limitations that previously hindered widespread adoption in applications requiring rapid, extensive information processing. The model's ability to outperform frontier models on tasks like web navigation suggests a new generation of AI agents that can operate more autonomously and efficiently, impacting industries reliant on automation and data analysis.

The implications extend to sectors requiring constant interaction and rapid feedback loops, such as education and customer support, where Mercury 3's near real-time responses can enhance user experience and operational effectiveness. For developers, its OpenAI API compatibility offers a low-friction integration path, potentially accelerating the deployment of advanced AI capabilities across a broader range of products and services without extensive refactoring.