Mercury 3: a diffusion LLM at 1,000+ tokens per second

Inception has released Mercury 3, which it calls the most capable diffusion language model released so far (Inception on X). Most language models write a response one token at a time, each step waiting on the last; Inception's diffusion models instead refine the whole output in parallel, as the company has described its earlier Mercury models. Inception credits that approach for the speed it advertises: over 1,000 tokens per second, which the launch blog puts at 1,068 tokens per second on NVIDIA GPUs (Inception blog).

The release is pitched as evidence that the approach scales. Inception says Mercury 3 has more than 40 times the parameters of Mercury 1 and scores three times higher across standard benchmarks by its own measure while running just as fast, and the context window grows to 1M tokens from 260K in Mercury 2.5. List pricing is $0.20 per million input tokens and $0.70 per million output tokens, halved during a launch discount, and the model is available through Inception's API, Baseten, OpenRouter and Vercel.

The headline comparison, that Mercury 3 outperforms OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 across a range of benchmarks, is Inception's own claim, presented in the company's own benchmark chart rather than in an independent evaluation. The nearest independent reference is Vals AI's run of the previous Mercury 2.5, which found it extremely fast but lower-scoring across its benchmark suite, so whether the new model closes that gap remains to be measured.