A Diffusion Model Writes Text at Record Speed

Most language models write the way a person reads aloud, one token after another, each word waiting on the one before it. Inception's new Mercury 3, released this week, takes a different route. It borrows the diffusion method that powers image generators, starting with a rough draft of the whole passage and refining it in parallel. The payoff the company is selling is speed.

The headline numbers

Inception says Mercury 3 reaches 1,068 tokens per second on Nvidia hardware. Its own comparison table puts Claude Haiku 5.5 at around 240 and OpenAI's GPT-6 Luna near 130, and the company claims Mercury 3 beats both on 8 of 12 benchmarks. The context window grows to one million tokens, up from 260,000 in the previous version, and the company says the model is more than 40 times larger than the first Mercury, though it did not disclose a parameter count.

Those results are self-reported, and the gap between a vendor's bench and a working desk is worth keeping in view. An independent coding test by the tooling firm Bito measured a median of 198 tokens per second for Mercury 3, faster than the roughly 92 it clocked for Claude Opus 5.5, but a long way short of the headline figure. One reviewer called the quality claims unverifiable until outside testing catches up.

Where it fits

Diffusion for text is still the unusual choice, and a credible model running at these speeds gives the approach its strongest showing yet. Mercury 3 is available through Inception's own API and through OpenRouter, Baseten and Vercel. The weights are closed, so this is a product to rent rather than a model to run yourself.

Speed has become its own battleground as the cost of the cheapest tier keeps falling, a race Anthropic joined when it put real reasoning into the budget-priced Haiku 5.5. Whether raw throughput translates into a better answer is a separate question, and one the benchmark numbers do not settle on their own.