NederlandsAI model Mercury 3 writes more than a thousand tokens per second by creating the whole text at once
The American company Inception has released Mercury 3, a language model that does not write word by word but brings a whole text into focus in a few steps. According to the company, this makes it four to eight times as fast as comparable models.
Almost all well-known AI chatbots write their answer the way a person types: piece by piece, from left to right. Each new token depends on everything that came before it. That works well, but it takes time. Especially for AI agents, which have to write something dozens of times for one job, those seconds add up.
The American company Inception is taking a different route. On 19 October 2026 it released Mercury 3. According to Inception, it is the most powerful language model ever to work in this different way.
A text that comes into focus, like a photo
Mercury is a so-called diffusion model. That technique is known from AI image generators. Such a model starts with noise and turns it into an ever sharper image in a number of steps.
Mercury does the same with text. The model first puts down a rough version of the whole answer and improves it in a few rounds, in all places at once. Compare it with a photographer who brings a whole photo into focus, instead of a painter who works line by line. Because a computer chip can do many things at the same time, this gives a lot of speed.
The figures according to Inception
According to Inception, Mercury 3 reaches about 1,068 tokens per second on Nvidia chips. A token is a piece of text, often a short word or part of a word. The company says this makes the model four to eight times as fast as the models it compares itself with, such as Claude Haiku 5.5 from Anthropic and GPT-6 Luna from OpenAI. Mercury 3 is also said to beat those models on a series of tests. Inception has not published the scores involved.
According to the company, the model is more than forty times as large as the first Mercury from 2025. It can take in an amount of text of one million tokens at once, enough for a few thick books. In a test in which the model had to navigate websites by itself, it completed 25 of the 50 tasks.
Mercury 3 can be used through Inception's own service and through platforms such as OpenRouter. Developers can try it free of charge up to 100 million tokens. After that it costs 20 US cents per million tokens you put in and 70 US cents per million that come out. At launch, half that price applies.
What still has to be shown
All performance figures come from Inception itself. Independent measurements still have to show how fast and how smart the model is in practice. For earlier Mercury models, the measuring firm Artificial Analysis did confirm speeds of more than a thousand tokens per second.
Inception was founded by researchers from Stanford University, among others. Co-founder and chief executive Stefano Ermon spent years researching diffusion models there.
What does this mean for the road to the singularity?
Most progress in AI comes from larger models and faster chips. Mercury shows that there is another way: with a smarter way of writing, you get more out of the same chips. That makes AI cheaper and faster, and therefore more useful.
That matters for AI agents. An agent carrying out a job thinks, looks something up, tries something and corrects itself, often dozens of times in a row. If each step goes four times as fast, a job of an hour becomes a job of a quarter of an hour. Whether diffusion will overtake the usual way of writing remains to be seen in the coming years. That there are now two strong routes side by side speeds up the search for the best one.
What is the technological singularity? Read the explanation →
