Mercury 2.5

Mercury 2.5 is a language model launched by Inception on 8 September 2026. It is a diffusion model, so instead of writing one token after another it produces and refines chunks of text in parallel. Inception says this is the source of its speed and aims it at developers who need low response times.

What Mercury 2.5 does

• Parallel generation: text is drafted and refined in blocks rather than token by token.
• Reasoning at speed: Inception describes it as the fastest reasoning model in production, a claim that has not been independently reproduced.
• Long context: a 260K-token context window.
• Multiple access routes: reachable via Inception's own API, on Baseten and on OpenRouter, and also as dedicated deployments for enterprises.

Clients

Inception cites Augment Code and OpenCall as customers, both reporting latency gains in the launch announcement; those results are vendor-reported.

Key facts

• Launched: 8 September 2026
• Maker: Inception
• Compared with: its predecessor Mercury 2, which Inception says it improves on, though no benchmark is named
• Not stated: licence and parameter count