Inception launches Mercury 3, a diffusion LLM that writes over 1,000 tokens per second

Inception released Mercury 3 on October 19, 2026 and calls it the most capable diffusion language model so far. Most chatbots write one token after another. Diffusion models like Mercury draft and refine many tokens at once, which is where the speed comes from.

According to Inception, Mercury 3 runs at 1,068 tokens per second on NVIDIA GPUs, roughly four to eight times faster than the speed-focused models it compares against, GPT-6 Luna and Claude Haiku 5.5, while matching or beating them on its chosen benchmarks. The model has more than 40 times the parameters of Mercury 1, and Inception says its benchmark intelligence has roughly tripled since that first mini version, evidence in its view that diffusion models scale like regular ones. The context window grows to 1 million tokens, up from 260,000 on Mercury 2.5.

List pricing is $0.20 per million input tokens and $0.70 per million output tokens, cut in half during the launch period, with cached input 90% cheaper. Early partners use it for browser agents, tutoring, search on smart glasses and customer support. Mercury 3 is available through Inception's own API, OpenRouter, Baseten and Vercel, and works with OpenAI-style API calls. Older Mercury 2 and 2.5 models are now closed to new users.

Why it matters

If diffusion models really keep up on quality, voice bots and multi-step agents that are slowed by token-by-token generation could get much faster and cheaper.

Source: Inception