Inception is an AI research and product company. We’re building diffusion-based language models for production applications.

Frontier LLMs today are autoregressive: they generate one token at a time, left to right. That design is inherently sequential, which makes latency and cost scale with every additional token - and every additional step in a workflow. As AI systems move from single-shot prompts to multi-step agents, output speed becomes a first-order constraint. The more the model thinks at inference time, the more important it is that it can still respond fast enough to ship in real products.

Diffusion is the modeling shift needed for a future where AI is always-on and operates at high volume. Diffusion models generate through a coarse-to-fine process: rather than committing token-by-token, they iteratively refine an output in parallel over a small number of steps. This enables revision during generation and a fundamentally different speed-cost curve.

The bigger idea is simple: AI shouldn’t behave like a one-way typewriter. It should behave more like an editor.

We’re scaling this approach with Mercury, the world’s first commercially available family of diffusion large language models (dLLMs). Mercury models seamlessly integrate into existing LLM workflows and deliver 5x greater speeds with best-in-class quality. Mercury enables a future of intelligent, responsive, and scalable AI applications.