Inception released Mercury 2.5 on 9 September 2026, a language model that generates text by refining a whole draft in parallel rather than writing it one word at a time. The company reports 1,107 tokens per second on widely available NVIDIA GPUs — roughly an order of magnitude above what comparable conventional models sustain — and calls it, to its knowledge, the largest diffusion language model ever trained. It is API-only, with a 260,000-token context window.
Key facts
- The headline number: 1,107 tokens per second on standard NVIDIA GPUs.
- When: announced 9 September 2026.
- Who: Inception, whose CEO is Stanford professor Stefano Ermon.
- Primary source: Inception's announcement.
Almost every model in production writes the same way: pick a word, append it, condition on everything so far, pick the next. That sequential dependency is the reason long outputs feel slow, and no amount of hardware removes it — token 900 cannot start until token 899 exists. It is also why inference is memory-bound rather than compute-bound for most deployments.
Diffusion language models attack the constraint at its root. Borrowing the machinery behind image generators, the model begins with a corrupted or masked version of the entire response and denoises it over a handful of passes, revising many positions simultaneously. The number of passes is roughly fixed, so generating a long answer costs far less than proportionally more time.
The analogy is the difference between a scribe and a sculptor. The scribe writes a letter one word at a time, and a longer letter takes proportionally longer. The sculptor starts with a rough block that already has the shape of the whole figure and refines the entire surface on each pass — a bigger statue does not require more passes, just more work per pass, which is precisely the kind of work a GPU is built to do in parallel.
Inception's claim is that this approach has now caught up on quality, not just speed. The company reports a 40% increase in what it calls intelligence over Mercury 2, and positions the result against cost-optimised frontier tiers — GPT-5.6 Luna at low reasoning, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. That is a deliberate and honest comparison class: Mercury 2.5 is pitched against the fast, cheap tier, not against frontier reasoning models. Alongside the throughput number, the company reports time-to-first-token under 170 milliseconds for its voice variant, which matters more than raw throughput for conversational use.
Pricing is $0.20 per million input tokens and $0.75 per million output, with a launch discount cutting both by 80%, to $0.04 and $0.15. Access is through the Inception API, Baseten and OpenRouter, with 100 million free tokens to start. Capabilities include tunable reasoning, parallel tool calls and schema-aligned JSON output — the practical feature set for agent work, where a model that emits structured output ten times faster changes what is affordable to build. NVIDIA's Shruti Koparkar and customers including OpenCall and Augment Code appear in the announcement.
Why it matters is narrower and more interesting than "a fast model shipped." Diffusion has been the perennial almost-ready alternative in language modelling for several years — theoretically appealing, practically behind. If a diffusion model is genuinely competitive with the cheap autoregressive tier while being several times faster, the architectural monoculture of the last five years has a real competitor for the first time. That is worth more than the benchmark position.
The honest caveats are substantial. The weights are closed, so nobody outside Inception can verify the throughput figure independently, and vendor-reported tokens per second are measured under conditions vendors choose — batch size, prompt length and hardware all move that number a great deal. "40% more intelligence" is a marketing construction rather than a measurement, and the comparison set is the budget tier by the company's own framing. Diffusion language models have historically traded away something for their parallelism, typically on tasks requiring strict left-to-right logical dependency, and the announcement does not dwell on where that trade still bites. Independent throughput measurements, not the vendor's, are the thing to wait for.
For readers tracking the economics rather than the architecture, this belongs with the broader collapse in inference cost that has defined the past year.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)