Mercury 3: diffusion model claims 1,000 tokens/s, beats GPT-6 Luna

Speed is the claim. Inception, a California AI research company, has released Mercury 3, and says it generates more than 1,000 tokens per second on widely available NVIDIA GPUs. The company says the model beats two speed-optimized rivals, GPT-6 Luna and Claude Haiku 5.5, on a range of benchmarks while running four to eight times faster.

Most large language models write one token after another, each one predicted from the last. Tokens are the chunks of text a model reads and writes. Diffusion language models work differently: they generate and refine several tokens in parallel. That is why they can answer faster, and it is also why developers have long had to choose between a small, quick model and a more capable, slower one. Inception argues Mercury 3 breaks that trade-off, keeping its speed while handling harder tasks.

The model is the fourth from a company that released its first version, Mercury 1, in June 2025. Mercury 2 added reasoning, and Mercury 2.5 added adjustable reasoning and expanded the context window to 260,000 tokens. Mercury 3 raises that window to 1 million tokens, and Inception says it is more than 40 times larger than Mercury 1. Stefano Ermon, co-founder of Inception, said the question customers asked most was whether diffusion could scale. The company says training of its next model has already begun.

Who gains, and when? Developers building agents, which chain several model calls together, should see the most immediate benefit, since each call returns faster. Education is an early case: the company reports that RevisionDojo, an edtech firm, cut response times for student feedback, including hints, definitions and marked answers, to about one second. Live translation and interactive tutoring are other targets the company names. Mercury 3 is already reachable through the Inception API, Baseten, OpenRouter and Vercel, and it uses the OpenAI API format, so existing applications can switch to it without extensive rewriting.

The caveats matter. The benchmark results and speed figures come from Inception itself, and the customer examples are early-access testers, not a broad survey of users. Speed in a controlled setup does not guarantee speed at scale, nor answer quality on every task. What will count is whether independent tests and everyday production use confirm the gains.