On October 19 Inception released Mercury 3, a diffusion language model. Inception says it generates 1,068 tokens a second on NVIDIA GPUs (their own measurement), with a 1M-token context.
For comparison, many chat models stream at somewhere in the tens to low hundreds of tokens a second. This is a different kind of speed.
Diffusion vs the usual way, in one paragraph
Most language models are autoregressive: they write one token, then the next, then the next, each depending on everything before it. You can't write token 500 until 499 exists.
A diffusion model starts with a rough draft of the whole output and refines all of it over a few passes, the way image models sharpen noise into a picture. Many tokens get worked on at once, so the time depends more on the number of passes than on the length of the answer.
That's the source of the speed. It's also why diffusion models were long thought to be worse at text, where order matters. Mercury 3 is a bet that this gap is closing.
What instant generation would change
If generation really is close to free in time, some designs that were too slow become reasonable:
1. Generate several, pick the best. Write five candidate answers and pick with a cheap check (tests pass, schema validates, a decision model scores it). That used to multiply latency by five.
2. Check everything before showing it. A validation step after generation is easier to justify when generation took 200ms instead of 8 seconds.
3. Agents with more, smaller steps. Agent loops are mostly waiting. Faster steps mean a 30-step task finishes while the user is still looking at the screen.
4. Autocomplete for things bigger than a line. Whole functions or whole SQL queries as you type.
What it doesn't change
- Speed isn't correctness. A wrong answer delivered in 50ms is still wrong. If anything, fast wrong answers are more dangerous, because they're easier to accept without reading.
- The rest of your pipeline. Network, retrieval, tool calls, database queries. If those take 2 seconds, a fast model saves less than you think.
- Your eval set. A new kind of model needs testing on your tasks more than a new version of a familiar one.
How I'd test it
Pick one feature where latency actually hurts users, not one where it just feels slow to you. Measure end-to-end time today, broken down: how much is the model, and how much is everything else? If the model is under half of it, a 10x faster model won't change much. If it's 90%, try it.
What I'd tell you before you start
Very fast generation changes which designs are affordable, not which answers are right. Use the speed to check more (more candidates, more validation, more tests), not to skip checking.
Source: Inception launch post (Oct 19, 2026), via "AI Model Releases: October 2026 Tracker", Digital Applied. Speed figures are the vendor's own.
Top comments (0)