Inception Labs

The diffusion category in this catalog is two rows, and both are this lab's: inception/mercury-2 and inception/mercury-2.5-preview are the only rows in the current OpenRouter snapshot whose listings describe a diffusion model. The descriptions Inception wrote for them define the class against itself — Mercury 2 is "the first reasoning diffusion LLM (dLLM)", Mercury 2.5 "the latest" — and no other listing in the snapshot claims one. A reader comparing this lab with any other is comparing a different mechanism, not a faster one of the same kind.

Inception's own pages make the pitch the router repeats. The about page describes diffusion-based large language models (dLLMs); it calls Mercury 'the world's first commercially available family of diffusion large language models'source, accessed 2026-09-03, and the mechanism is the pitch: diffusion rather than autoregressive generation: models produce many tokens in parallel, which Inception says makes them 'several times faster and less than half the cost of conventional LLMs'source, accessed 2026-09-03. The team is leading researchers and engineers from Stanford, UCLA, Cornell, Google DeepMind, Meta AI, Microsoft AI, and OpenAIsource, accessed 2026-09-03, and the lab says it is currently deploying these diffusion LLMs at Fortune 500 companiessource, accessed 2026-09-03 — all of it the vendor speaking. The catalog's part of the story is narrower: two rows, and the newer one is an explicit preview — API-only, no weights released, per its llm-releases entry.

Six months apart — 4 March to 31 August 2026 — the second was announced as a preview positioned above the first. The about page links the papers behind the mechanism, under "some of the technologies we've developed": among them, the foundation line (Diffusion Models, Flash Attention, Direct Preference Optimization) and the discrete-diffusion line Mercury descends from (Masked Diffusion, Block Diffusion, Remasking Diffusion, and d1 Reasoning, the April 2025 framework that adapts pre-trained masked dLLMs into reasoning models via SFT and RL). The family page footnotes that Mercury 1 and Mercury Edit 2 "remain supported for existing customers" — the only commitment a listed model carries there.