Diffusion-based generation
The model generates through a diffusion-based coarse-to-fine process rather than committing to output one token at a time, supporting fast inference for production workflows.
Mercury 3 is Inception’s flagship diffusion-based large language model for production applications that need fast responses, reasoning, and long-context processing. It supports OpenAI-compatible workflows and is aimed at browser agents, education, real-time search, customer support, and other latency-sensitive systems.
Mercury 3 is a diffusion-based large language model from Inception for production applications that need both reasoning and low-latency responses. It is the flagship model in Inception’s Mercury portfolio and is positioned for multi-step systems where waiting on model output can compound across agent steps.
The model provides a 1-million-token context window and reports 1,068 tokens per second on NVIDIA GPUs. It is available through Inception’s API Platform and several external access routes, with an OpenAI-compatible API intended to fit existing LLM applications. Use cases described by Inception include browser agents, interactive education, real-time search, and customer support.
The model generates through a diffusion-based coarse-to-fine process rather than committing to output one token at a time, supporting fast inference for production workflows.
The published benchmark figure is 1,068 tokens per second on NVIDIA GPUs, with the site describing performance above 1,000 tokens per second for commercial GPU deployments.
Mercury 3 supports a 1-million-token context window, allowing applications to provide substantially larger inputs than the 260,000-token context reported for Mercury 2.5.
The model is OpenAI API compatible and is supported through AISuite, LiteLLM, and LangChain, reducing changes required to connect an existing application.
The model is available through the Inception API Platform and third-party access routes including OpenRouter, Baseten, and Vercel. Enterprise routes include AWS Bedrock and Azure Foundry.
Enterprise deployments can use dedicated capacity, autoscaling, configurable data retention, compliance controls, and service-level agreements, subject to the selected arrangement.
Browser-use agents can call the model at each navigation, research, or log-analysis step. Faster responses reduce accumulated waiting across multi-step agent workflows.
Tutoring and education products can use the model for conversational voice assistance, grading, hints, definitions, and adaptive responses when students need immediate feedback.
Search systems can process documents and return answers while a user remains engaged. Inception cites applications including real-time search agents and smart-glasses experiences.
Support systems can use Mercury 3 for conversational responses across chat, voice, and email, including workflows where response accuracy and task completion are important.
Teams building production AI applications can connect through the OpenAI-compatible API or use enterprise deployment routes when they need capacity, scaling, governance, or retention controls.
Mercury 3 is a diffusion-based large language model designed for production applications that need fast responses and reasoning capability.
It is available through the Inception API Platform and OpenRouter, Baseten, and Vercel. Enterprise deployment options also include AWS Bedrock and Azure Foundry, while model routers include OpenRouter and Models.dev.
Mercury 3 is OpenAI API compatible and can be used through libraries including AISuite, LiteLLM, and LangChain. The source also describes it as a drop-in replacement for traditional LLM workflows.
The published standard price is $0.20 per million input tokens and $0.70 per million output tokens. At launch, the site lists a 50% discount, making those rates $0.10 and $0.35 respectively; cached input tokens are listed at $0.02 per million, or $0.01 at the launch discount.
Enterprise deployments support dedicated capacity, autoscaling, compliance controls, configurable data retention, service-level agreements, and support options. The enterprise page states that prompts and outputs are treated as customer data and are not used for training.