Mercury 3 is Inception's newest language model. Instead of writing its answer one token after another, it improves rough text step by step into a final reply, a method called diffusion.
In plain words
Mercury 3 does not build a reply one token at a time. As a diffusion model, it starts from noisy text and refines it over several passes into finished blocks. Inception says it beats Claude Haiku 5.5 and GPT-6 Luna on a range of tests while running 4 to 8 times faster, producing 1,068 tokens a second on NVIDIA GPUs. It can take in up to 1 million tokens at once. Developers reach it through Inception's API, which works like OpenAI's, or through OpenRouter, Baseten and Vercel.
A quick example
Imagine a homework helper you talk to out loud. You ask why your answer to a maths question is wrong, and a long pause would break the conversation. A fast model lets the helper think through your working and reply while you are still paying attention.
Where you'll meet it
- Voice helpers, tutors and support chats that must answer without a pause.
- Web agents that call the model at every step of a task.
Related
Sources
- DocsDiffusion at Scale: Introducing Mercury 3, Inception · read 19 Oct 2026
- DocsStreaming & Diffusion, Inception Platform docs · read 19 Oct 2026
Checked by WhatIsAI against the sources above · 19 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.