Inception launches Mercury 3 diffusion AI model, claims 4-8 times faster speed

Inception, a US artificial intelligence start-up, released Mercury 3 on October 19, 2026, a large language model built on a technique called diffusion rather than the word-by-word method used by most chatbots. The company says the model generates 1,068 tokens per second on Nvidia graphics chips, runs four to eight times faster than comparable models, and can handle a context of one million tokens. It is available through Inception's own platform and services such as OpenRouter. Independent experts have noted that the launch materials do not yet include benchmark scores that can be checked against standard tests.

Most AI chatbots work like a very fast typist: they produce text one token at a time, each word depending on the ones before it. Diffusion models work differently. They start with a rough, noisy draft of the whole answer and refine it in a few passes, improving many words at once. The same idea powers popular image generators. Its supporters argue that for text it can be much faster and cheaper, because the computer can work on many parts of the answer in parallel.

Speed matters more than it might seem. For coding assistants, customer service bots or voice agents, a delay of a second or two can make a product feel slow. Faster models can also be cheaper to run, which makes AI more affordable for start-ups and for services aimed at price-sensitive users, including in India.

But speed is only useful if the answers are good. That is why the lack of independently checkable benchmark scores matters. Companies often present results on tests they choose, under conditions they control. Until independent evaluators run the model on standard benchmarks, buyers should treat the performance claims as unverified and test the model on their own tasks.

For Indian developers and AI start-ups, Mercury 3 is another option in a crowded field. Because it is available through common platforms, teams can try it alongside other models. Its long context window could help with tasks such as reading long legal or financial documents, and its speed could suit real-time applications. Whether it is good enough in Indian languages is something users will need to check for themselves; the launch materials do not make specific claims about them.

Short-term: Developers can test Mercury 3 through existing platforms, and independent benchmarking groups are likely to publish their own scores in the coming days and weeks.

Long-term: If diffusion models prove they can match conventional models on quality while being much faster, they could push the industry towards cheaper, quicker AI services. If not, they may remain a niche choice for tasks where speed matters more than depth.

Who is affected: AI developers and start-ups choosing which models to build on, companies running chatbots and coding tools, cloud and chip providers, and users of AI products who notice speed and cost.

Questions and Answers