Mercury 3 generates more than 1,000 tokens per second

What this is about

Inception released the Mercury 3 language model on October 19, 2026. The company says it generates 1,068 tokens per second on Nvidia GPUs, making it four to eight times faster than GPT-6 Luna and Claude Haiku 5.5. The key difference is not simply more computing power: Mercury 3 is a diffusion model and produces text differently from widely used autoregressive models.

This matters most where every pause is noticeable. Voice assistants, search services, learning software, and browser agents make many model calls in sequence. Even a few hundred milliseconds per step can add up to a noticeable wait.

What Mercury 3 actually does

Conventional language models generate output one token at a time. Each new piece of text depends on the preceding one. Mercury 3 instead starts with a rough draft across multiple text positions and improves it over several passes. This lets the model work on several parts of an answer at once.

Inception specifies a one-million-token context window. Regular API pricing is $0.20 per million input tokens and $0.70 per million output tokens; the vendor says launch prices are half those amounts. The API is compatible with the OpenAI format and is also available through OpenRouter, Baseten, and Vercel.

Mercury 3 won eight of twelve comparisons in Inception's own tests. It scored 76.84 on IFBench and 51.08 on SWE-Bench Atlas. The picture is not uniform, however: Claude Haiku 5.5 led on the longer DeepSWE coding test and on retrieval from long texts.

Why it matters

Speed changes which applications are practical. An agent operating a website in 30 steps waits for the model 30 times. Reducing generation time at every step does more than make the interface feel smoother; it can also give the agent time to try alternative approaches.

One early Inception partner tested Mercury 3 on 50 live navigation tasks and reported a 50 percent completion rate. A provider of search tools for smart glasses reported median model generation of 0.9 seconds and a total response time of 2.2 seconds. These examples show possible value, but they are not independent comparative studies.

The release also matters technically. Diffusion models are familiar from image generation, but research such as Google's DiffusionGemma explores the same approach for text. Mercury 3 is now a commercial example suggesting that this architecture can scale to larger models.

In plain language

An autoregressive model works like someone writing a letter with a ballpoint pen: one word after another, with no easy way to jump back to the beginning. A diffusion model is more like a person using a pencil to sketch across the whole page first and then improve several areas at once. That can be faster as long as the revisions fit together reliably.

A practical example

Suppose a support agent must search ten documents, check three rules, and formulate a reply. With 13 model steps and two seconds of generation per step, generation alone adds 26 seconds of waiting. If a faster model completed the same step in half a second, that portion would fall to 6.5 seconds.

This is only an illustrative calculation. Real systems also include network delays, tools, database access, and possible retries. High token speed is of little benefit if the model makes wrong decisions more often.

Scope and limits

First, most speed, quality, and partner results come from Inception. Independent tests on identical hardware and under realistic traffic spikes are not yet available. Second, the company does not disclose the exact model size, limiting comparisons of efficiency and resource needs.

Third, a high token rate does not automatically mean a better answer. The published tables show mixed results for coding, long documents, and demanding reasoning. Organizations should test their own tasks, error rates, end-to-end latency, and cost. The introductory discount also has no clearly stated end date and should not be treated as a permanent basis for budgeting.

SEO & GEO keywords

Mercury 3, Inception Labs, diffusion model, diffusion LLM, language model, tokens per second, AI agents, OpenAI-compatible API, text generation, Nvidia GPU, model latency, DiffusionGemma