Inception has introduced Mercury 3, a diffusion-based large language model (LLM) designed to provide high-speed inference for complex commercial applications. Unlike standard autoregressive models that generate text one token at a time, diffusion models generate content by iteratively refining a noise-based representation. Mercury 3 is positioned as a scalable evolution of this architecture, with the developers claiming it outperforms frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5 while delivering speeds of over 1,000 tokens per second on NVIDIA GPUs.

The model represents a significant scale increase from its predecessors, featuring more than 40x the total parameters of Mercury 1. According to Inception, this scaling has resulted in a 3x increase in intelligence across standard benchmarks compared to the Mercury 1 (mini) version. The technical goal of Mercury 3 is to prove that the scaling laws governing autoregressive models—where increasing parameters and data leads to higher intelligence—also apply to diffusion-based architectures. This allows the model to handle sophisticated tasks such as web navigation, real-time tutoring, and data engineering without the latency bottlenecks typical of larger autoregressive systems.

A chart comparing model quality against model size for the Mercury series.
A visualization of the scaling laws showing the relationship between model size and intelligence for Mercury 3.www.inceptionlabs.ai

The chart illustrates the scaling laws and the 3x increase in intelligence achieved by increasing the parameter count.

In practical production environments, Mercury 3 is being utilized for latency-sensitive 'browser use' agents, where small delays in model response can compound into significant wait times for the user. In testing against the Mind2Web navigation benchmark, Mercury 3 completed 50% of tasks, outperforming GPT-6.1-Sol while using the same token volume and number of agent steps. Additionally, the model's speed enables real-time interactive education, where it can provide instant feedback and hints to students in approximately one second. While the model offers a high-speed alternative for these use cases, it is currently available via the Inception API Platform and several third-party providers like OpenRouter and Vercel.

A benchmark comparison chart for Mercury 3 performance.
Benchmark results demonstrating Mercury 3's performance in navigation and speed tasks.www.inceptionlabs.ai

This figure provides a visual comparison of Mercury 3's performance against other models like GPT-6.1-Sol.