Mercury

Mercury writes tokens in parallel rather than one at a time, the core reason it's several times faster than conventional models. Inception reports a sub-300ms time to first token and 1,107 tokens per second on widely available NVIDIA GPUs. That gap sounds technical until you use it. For a user, that's the difference between an assistant that feels instant and one that makes them wait while a spinner turns, which is exactly the kind of small delay that quietly pushes people away from a product.

The Mercury 2.5 model accepts up to 260,000 tokens of context in a single request. That's enough to hold long documents, big code files, or a lengthy conversation history without chopping it up. Long context plus high speed is an unusual pairing, since most fast models keep their windows modest, and teams that work with large codebases or lengthy reports tend to feel that gap immediately. Huge deal for big files.

You can dial how much internal reasoning the model does per request. Turn it up for hard problems that need step-by-step thinking; turn it down for simple lookups where extra reasoning just adds latency and cost. That control is rare in hosted models, where reasoning is usually on or off. Nice.

Mercury can constrain its output to match a JSON schema you define. Diffusion lets it fix up tokens as a group, which Inception says makes structured output more dependable. Fewer broken replies. Developers get less malformed output to clean up when they feed results into the next step of a pipeline, which is the kind of papercut that adds up fast once a workflow runs thousands of times a day in production.

The model can fire off multiple tool or function calls at once instead of waiting for each to finish. In an agent that checks a calendar, queries a database, and searches the web, that parallelism trims real seconds off the response. It fits the multi-step workflows Mercury is pitched for, and anyone who has watched an agent grind through five sequential calls knows how much that wait adds up over a single conversation. Big win for agents.

Mercury is a diffusion LLM, or dLLM. Inception describes it as among the largest diffusion language models trained to date. The architecture also opens a path to mixing text with audio, images, and video in one framework. Today's shipped product is text-first.