An interactive agent can lose its usefulness while waiting between steps. Browser navigation, tutoring and support combine model calls with tools and network delays, so generation speed matters most when it improves the complete task without increasing mistakes or retries.
Inception’s October 19 Mercury 3 announcement describes a diffusion-based language model scaled substantially beyond its first Mercury release. The company reports generation above 1,000 tokens per second, a one-million-token context window and availability through its API and distribution partners. Its comparisons and customer accounts are evidence supplied by the developer and its launch participants, not an independent replication of performance across applications. Read the primary source.
Speed is more than output rate
A system can produce a long answer quickly while still taking too long to begin a useful response. Conversely, a short action-selection call may spend more time in tool execution than in generation. Evaluators should therefore distinguish time to first useful output, completion time, tool latency and the number of attempts required to finish.
The release gives teams a model they can try in an existing compatible interface. A controlled comparison would hold the tasks, tools and stopping rules fixed, measure successful completions under a common budget, and inspect failures. Long-context availability should also be tested with the actual evidence distribution; a context capacity is not a guarantee that every relevant detail will be used correctly.
If diffusion preserves quality while reducing repeated waits, it could change the economics of interactive agents. The immediate decision is narrower: whether Mercury 3 improves the latency and reliability combination a particular product needs. Reproducible application-level results will be more informative than multiplying a generation-speed claim by the number of agent steps.
References
Questions, news tips or corrections? Contact [email protected].
