Inception builds large language models (LLMs) based on diffusion rather than the traditional autoregressive method. Its models, called dLLMs, generate many tokens at once instead of one at a time. The company says this makes them much faster and more efficient than conventional LLMs, and its flagship Mercury line is sold through an API.
Most LLMs write text one token at a time, so response speed is limited by that sequence. Mercury generates tokens in parallel, which the page says raises throughput and makes better use of GPUs. The company reports sub-300ms time to first token and 5-7x higher throughput for Mercury 3, and it says the models match the quality of speed-optimised frontier models.
The product is aimed at businesses building AI applications where response time matters, such as coding tools, voice agents, customer support, translation, and search over internal knowledge bases. The Mercury models are OpenAI API compatible, so developers can use them as a drop-in replacement for existing LLM calls. Libraries including AISuite, LiteLLM and LangChain are listed as supported.
The lineup includes Mercury 3 for complex reasoning, Mercury Voice for voice agents with time to first token under 170ms, and Mercury Decide for structured decisions with calibrated probabilities and escalation to a human.
Pricing is usage-based and charged per million tokens. Mercury 3 is listed at $0.10 input and $0.35 output per 1M tokens, with a 50% discount shown on the page. New API keys include 100 million free tokens. An Enterprise plan offers custom rate limits, SLA guarantees, security and privacy terms, and volume pricing. The models are also available through AWS Bedrock and Azure Foundry, and the company offers fine-tuning and private deployments on request.
