Move over, "fancy autocomplete." Inception Labs CEO Stefano Ermon joins The New Stack to explain how Mercury 2 uses diffusion to deliver LLM speeds 10x faster.

The New Stack is a publication covering trends and technologies in cloud-native development, DevOps, and software delivery. Developers can learn about containerization, Kubernetes, and cloud computing, as well as explore topics such as microservices architecture, serverless computing, and continuous integration/continuous delivery (CI/CD) pipelines.

The New Stack

Inception Labs has launched Mercury 2, a large language model built on diffusion architecture rather than the autoregressive approach used by OpenAI, Anthropic, and Google. Instead of generating text one token at a time, Mercury 2 refines answers in parallel — similar to how image diffusion models work — achieving over 1,000 tokens per second, which Inception claims is 5–10x faster than speed-optimized frontier models. Quality-wise, it matches Claude Haiku and Google Flash-class models rather than top-tier models. CEO Stefano Ermon, who pioneered diffusion models at Stanford and won Best Paper at ICML 2024, notes that diffusion architectures are naturally better suited for GPU parallel computation and reinforcement learning-based reasoning. Nvidia is an investor and is helping optimize the serving engine. Mercury 2 is available via an OpenAI-compatible API, with AWS Bedrock integration coming soon. Inception Labs is currently the only company shipping a production diffusion LLM.

Inception Labs says its diffusion LLM is 10x faster than Claude, ChatGPT, Gemini