Inception releases Mercury 3, a diffusion language model that generates more than 1,000 tokens per second

Inception, co-founded by Stanford professor Stefano Ermon, announced its diffusion-based language model Mercury 3 on October 19. Conventional language models write text left to right, one token at a time. Diffusion models instead refine a whole response in parallel over a few steps. According to the company, Mercury 3 runs at 1,068 tokens per second on NVIDIA GPUs. Its context window is 1 million tokens, up from 260,000 for Mercury 2.5. Inception says the model has more than 40 times the parameters of Mercury 1 and scores three times higher on standard benchmarks.

The company claims Mercury 3 outperforms OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 across a range of benchmarks while running 4 to 8 times faster. The announcement gives no per-benchmark scores. On X, Ermon called it the most intelligent diffusion LLM ever released. Shuhao Zhang, co-founder of early-access partner TinyFish AI, says the model completed 50% of 50 live Mind2Web web navigation tasks, significantly more than GPT-6.1-Sol. None of these results have been independently verified.

List price is $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount brings that to $0.10 and $0.35, with no announced end date. The model is available through the Inception API, OpenRouter, Baseten and Vercel, and is compatible with the OpenAI API. Inception’s models page says Mercury 2 and 2.5 are closed to new users, and existing customers keep access until November 15.