Overview

Mercury 3 is a diffusion‐based large language model from Inception Labs that pushes the upper limits of speed and context size. The model is marketed as surpassing GPT‑6 Luna and Claude Haiku 5.5 on a range of benchmarks while delivering 1 000+ tokens per second. It is 40× larger in total parameters than the original Mercury 1, yet remains fast and can process over 1 million tokens of context.

Typical users are developers building latency‑sensitive tools: real‑time voice assistants, interactive tutoring systems, web‑navigation agents, and search or support chatbots. In each scenario the model receives a prompt that can be up to 1 million tokens—such as a full web page, a transcript of a dialogue, or a batch of documents—and returns a completion that typically stays within the same token count as the input. In practice, a tutor bot can analyze a student’s typed answer, generate hints, and respond in about a second, while a search agent can parse dozens of passages and return a concise answer to a user within 2 seconds. Access partners have seen the model complete 50 percent of 50 live navigation tasks against frontier models, using equal token budgets and a similar number of agent steps but at a fraction of the per‑token cost.

The pricing model offers $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount of 50 percent ($0.10 input, $0.35 output). Cached tokens cost 90 percent less, which is useful for maintaining session‑level context across repeated calls. The listing does not disclose the API endpoints, SDKs, or supported host platforms—only mentions that the model runs on NVIDIA GPUs. It also does not specify any usage limits beyond the token rates or provide evidence of billing or license terms for production environments.

Because the description concentrates on speed, context, and cost, it is most suitable for applications that need to pull long documents into a single inference and respond quickly, such as AI tutors and search agents. It may be less relevant for workloads that require ultra‑low network round‑trip times, as there is no mention of deployment in edge or on‑premise environments, nor of specific latency targets beyond the per‑token rate.