Mercury 3, a diffusion-based large language model from Inception, is claimed to run at 1,068 tokens per second with a 1-million-token context window. The company announced the model in a post that highlights its speed and long-context capability.
Inception also published a benchmark table comparing Mercury 3 with GPT-6 Luna and Claude Haiku 5.5. According to that table, Mercury 3 leads on 8 of the 12 benchmarks. The results are from Inception's own evaluation, so they have not been independently confirmed.
The announcement does not include details on hardware, cost, or availability. As with any self-reported performance, the figures should be treated with caution until third-party testing is done.