Inception has released Mercury 3 through its API and several hosting platforms. The diffusion language model expands the announced context window to one million tokens compared with the previous version, targeting applications with long instructions and tool histories. Early users include web-agent, education and customer-service companies. Access uses the OpenAI-compatible request format; the release does not announce downloadable weights.
Inception, a developer of diffusion language models, released Mercury 3 on 19 October. The model is available through the company’s application programming interface, or API, and through OpenRouter, Baseten and Vercel, platforms that provide access to AI models. Its intended uses include agents that carry out tasks on websites, customer support and interactive education.[1], [2]
The announced context window grows to one million tokens, compared with the smaller Mercury 2.5 window. Tokens are the pieces of text a model processes; the window holds instructions, documents and the history of interactions with tools within a request. Developers can use the OpenAI-compatible request format when connecting their existing applications to Mercury 3.[1]
Early users include TinyFish, which builds web agents, General Learning’s RevisionDojo education product and Lorikeet’s customer-service workflows. Their experiences are customer accounts supplied in the announcement. Inception offers enterprise options including dedicated capacity, automatic scaling and configurable data retention. The launch provides hosted model access; it does not announce downloadable Mercury 3 weights.[1]
