Inception: Mercury 3

Inception: Mercury 3 is a text model from Inception, the lab behind the Mercury line of diffusion-based language models. Inception describes Mercury 3 as its most capable model to date, and positions it for agentic workflows and real-time applications where latency compounds. It is a text-only model in the sense that it accepts text and returns text, so it does not handle images, audio or other input types.

The context window is 1,000,000 tokens, which is large enough to hold entire codebases, long document sets or extended conversation histories in a single request. Output is capped at 131,072 tokens, so a single response can be very long even though the input budget is far larger. The model supports reasoning controls through the reasoning, reasoning_effort and include_reasoning parameters, and it can return structured output via response_format and structured_outputs. Tool use is supported through the tools and tool_choice parameters, which is what makes it a candidate for agent loops rather than plain chat.

Those capabilities point to a few practical patterns. Agentic workflows are the obvious one: you can give the model a large working context, expose tools, and let it decide when to call them, with reasoning effort tuned to the difficulty of the step. Real-time applications are the other stated fit, where the diffusion approach is meant to keep latency low as interactions accumulate. Long-context tasks such as reviewing a large repository, summarizing a document collection or maintaining a long-running assistant session also follow from the 1M token window, and structured outputs make it usable for extraction and classification pipelines that need predictable JSON.

You call it through api14's OpenAI-compatible API using the model id "inception/mercury-3". Because the surface is OpenAI-compatible, existing clients and SDKs work with a base URL change. The request parameters it supports are include_reasoning, max_tokens, reasoning, reasoning_effort, response_format, stop, structured_outputs, temperature, tool_choice and tools. Use max_tokens to bound the response length, temperature for sampling control, stop for custom stop sequences, and the reasoning and tool parameters when you want the model to think through a step or act on external functions.

One limit worth noting is modality: the model is text in and text out, so anything involving images or audio needs a different model. The provider release date is 2026-10-16. If your task does not need reasoning, tool calls or a very large context, a simpler text model may be a better fit, but for agentic and long-context work within the supported parameter set, Mercury 3 is the direct option on api14.