Mercury 2 (description https://www.inceptionlabs.ai/blog/introducing-mercury-2 : and demo: https://chat.inceptionlabs.ai/) is the first model for text generation based on delivery, as are image or video generation.

Instead of proceeding token by token and “predicting” the next game as current LLMs do, Mercury 2 proceeds by starting with “noise” and gradually refining its response.
This method makes the response faster (the processing speed comes from the model itself and not from the hardware used), even with the so-called “reasoned” “high” model, is less expensive and would greatly reduce “hallucinations”.

Four variations of the model are present: Instant, low, medium, high.

A “diffusion effect” button stages the broadcasting principle (see video capture).

During testing, web search did not work and the knowledge base was stopped until January 2025.

The demo version does not seem to have any limits, a window inviting you to pre-register for the future paid version appears every 5 interactions but can be closed.