Inception Labs, the company behind the Mercury diffusion LLMs

A startup born from diffusion research

Inception, often called Inception Labs, was founded by three US academics: Stefano Ermon (Stanford), Aditya Grover (UCLA) and Volodymyr Kuleshov (Cornell). According to its November 2025 funding announcement, the company is based in Palo Alto, California; its official pages do not state a founding year. Its backers include well-known names in the field such as Andrej Karpathy and Andrew Ng.

Its bet is a technical one. Nearly all of today’s large language models are “autoregressive”: they write one token at a time. Inception applies the logic of diffusion models, made popular by image generation, to text. The answer first appears as a noisy sketch, then sharpens over a few passes, with several parts processed in parallel.

The Mercury family

Mercury, launched in February 2025, is described by Inception as the first commercially available family of diffusion language models. As of October 22, 2026, the lineup lists three models:

• Mercury 3, the most capable general-purpose model in the family, with reasoning;
• Mercury Voice, tuned for voice agents, where the delay before the first word is critical;
• Mercury Decide, a decision model built for structured choices such as classifying or routing requests.

What they share is a focus on latency: the company highlights very short response times and a lower cost per task than speed-optimized models. These are Inception’s own claims; to compare models on measured criteria, use our model comparison tool.

Closed models, API access

Inception does not release Mercury’s weights. These are proprietary models, the opposite of open-weight models. You can reach them in several ways:

• Inception’s API platform, documented at docs.inceptionlabs.ai;
• cloud marketplaces such as AWS Bedrock and Azure Foundry;
• aggregators such as OpenRouter;
• Mercury Chat, a browser-based chat interface for trying the models.

What sets Inception apart

In practice, Inception has carved out a niche: speed. While most labs compete on the raw intelligence of ever-larger models, Inception relies on a different architecture to serve real-time uses such as voice or AI agents that chain many calls together. One open question remains: diffusion is still a minority approach for text, and how it really compares with the best autoregressive models depends on the task. To see where Inception fits among other players, browse our list of AI labs.