Mercury 2.5 LLM achieves 770 tokens per second

A language model designated Mercury 2.5 has been measured at 770 tokens per second, according to a report circulated on Hacker News on September 23, 2026, and linked to a listing on the Artificial Analysis models page.

The supplied material provides no elaboration on the model’s architecture, its developer, or the conditions under which the throughput figure was obtained. The Artificial Analysis page is cited as the source of the metric, but the material available does not describe the methodology behind the measurement, nor does it specify what other models, if any, were tested alongside Mercury 2.5.

The Hacker News discussion is the only reported venue for commentary on the figure. The material does not indicate the content or tenor of that discussion beyond the headline claim itself.

No further technical detail, comparative benchmark, or statement from a developer accompanies the report. The figure of 770 tokens per second stands as the sole reported data point regarding Mercury 2.5’s performance in this material.

Readers seeking architectural explanation, provenance of the model, or context for how this throughput compares to other language models will find no such information in the sources provided. The Specola records the figure as reported and nothing beyond it.