Mercury 2: how a diffusion language model hits 1,000 tokens per second

If you work with AI regularly, you know the feeling: you hit enter and watch the answer appear word by word, as if an invisible typewriter were at work. For an email, that is fine. For a voice assistant, or an agent that takes twenty steps in a row, that waiting time is the whole problem.

On 24 February 2026, the US company Inception released Mercury 2, a language model that skips the typewriter. It does not write token by token; it refines an entire answer at once. According to Inception, that gets it to more than 1,000 tokens per second, over ten times the speed of the fast models from Anthropic and OpenAI.

What is a diffusion language model?

A diffusion language model produces text by refining a rough draft of the whole answer over a number of steps, instead of predicting it word by word. Each step changes many positions at once. The same principle drives image and video generators such as Stable Diffusion and Sora: they start from noise and turn it into an image step by step.

The well-known language models, such as the GPT series, Claude and Gemini, are autoregressive. They predict one token, append it to the text, then predict the next. That is sequential by definition: token 500 cannot be computed until token 499 exists, no matter how many GPUs you throw at it.

Think of the difference between a typewriter and an editor. The typewriter puts one letter after another on paper and cannot go back. The editor looks at the whole text and fixes several places at once. A diffusion model can still change something earlier in the answer during a later refinement step, and Inception presents that as built-in error correction. How much that matters in practice is hard to say: reasoning models also correct themselves, they just do it in their thinking steps, before the actual answer.

How fast is Mercury 2?

Inception measures 1,009 tokens per second on NVIDIA Blackwell GPUs. For comparison, the company cites about 89 tokens per second for Claude Haiku 4.5 (with reasoning) and 71 for GPT-5 mini. A 500-token answer takes Mercury 2 half a second. At 80 tokens per second, the same answer takes more than six.

What users feel is the total wait, including the thinking. There the gap is even larger. In the launch figures, as reported by The Decoder, a task takes Mercury 2 1.7 seconds end to end. Gemini 3 Flash needs 14.4 seconds, GPT-5 mini 22.8 and Claude Haiku 4.5 23.4.

These are Inception's own measurements, on its own hardware and settings. Independent comparisons may come out differently.

Is Mercury 2 smart, or just fast?

Speed is worthless if the answer is wrong. Inception calls Mercury 2 the first diffusion language model that can reason, with an adjustable amount of thinking. The company published these scores:

• AIME (maths competition): 91
• GPQA Diamond (graduate-level science): 74
• IFBench (following instructions precisely): 71
• LiveCodeBench (coding): 67

That puts Mercury 2 among the fast, cheap models, not alongside today's heaviest models. Inception does not claim otherwise: it positions the quality as comparable to the best speed-optimised models. The promise is not "smarter" but "just as smart, and much faster".

Where does the speed make a difference?

In a chat window, the difference between two and six seconds is pleasant but not decisive. It gets interesting where waiting times pile up.

• Voice. In a conversation, a silence of a few seconds is noticeable straight away. In July 2026 Inception released a separate voice version of Mercury 2, with under 300 milliseconds to the first token, according to the company.
• Agents and RAG. An agent that plans, searches, calls a tool and plans again waits for the model at every step. Ten steps of five seconds is almost a minute; ten steps of half a second you barely notice. The same goes for a RAG pipeline that asks a model several times per question which documents are relevant.
• Coding. Code that appears almost instantly keeps you in your flow. For small edits and completions, speed matters more than the very best reasoning.

Who is behind it, and what does it cost?

Inception was founded by three professors who did research on diffusion models themselves: Stefano Ermon (Stanford), Aditya Grover (UCLA) and Volodymyr Kuleshov (Cornell). In November 2025 the company raised 50 million dollars in a round led by Menlo Ventures, with the venture arms of Microsoft (M12) and Nvidia (NVentures) among others. Andrew Ng and Andrej Karpathy invested personally.

Mercury 2 is available through an OpenAI-compatible API. Existing code can in principle keep working if you change the base URL and the model name. The context window is 128,000 tokens. The price: 0.25 dollars per million input tokens and 0.75 dollars per million output tokens. For comparison, Inception cites 0.50 and 3.00 dollars for Gemini 3 Flash.

Has the latency wall been broken?

For one kind of work, yes: fast, cheap calls where waiting time matters more than the last few percent of quality. For the heaviest reasoning, a large autoregressive model remains the logical choice, and Inception more or less says so itself. Meanwhile the company keeps shipping: Mercury 2.5 came out on 8 September 2026.

The quickest way to feel what 1,000 tokens per second means is to try it in Inception's chat. Set the reasoning level high, give it a coding task and watch how fast the code appears. Chances are the typewriter will feel slower than you remember.

Further reading on this site

• OpenAI WebSockets: why agents with many tool calls get faster
• Programmatic tool calling: why AI agents increasingly call their tools through code
• Which AI model can you run on your own laptop? Measured on 16 GB, October 2026

Sources

• Inception: Introducing Mercury 2
• The Decoder: Inception launches Mercury 2, the first diffusion-based language reasoning model
• TechCrunch: Inception raises $50 million to build diffusion models for code and text
• Menlo Ventures: From the Lab to the Frontier, the story behind Inception
• Inception blog (Mercury 2 Voice, Mercury 2.5)