Mercury 2.5
Inception Labs' diffusion language model, released September 2026, which generates around 1,107 tokens per second on standard Nvidia GPUs.
Mercury 2.5 is the model Inception Labs uses to argue that diffusion language models are ready for everyday work. It runs at roughly 1,107 tokens per second on widely available Nvidia hardware, holds 260,000 tokens of context, and supports tool calls, schema-controlled JSON output, and tunable reasoning effort.
Inception positions its quality alongside the cost-optimised tiers of the big labs rather than their flagships, and prices it at 0.20 dollars per million input tokens and 0.75 per million output. Inception describes it as the largest diffusion language model trained so far, a claim no independent party has verified.