Inception Labs Launches Mercury 2 AI, Outpaces Google's DiffusionGemma

On June 21, 2026, Inception Labs introduced its Mercury 2 AI model, claiming it is the fastest reasoning language model available. Mercury 2 generates about 1,000 tokens per second, significantly outpacing competitors like Google’s DiffusionGemma, which produces similar performance but scored only 69.1% in benchmarks, compared to Mercury 2's score of 90% on AIME 2026, a mathematics assessment.

Both Mercury 2 and DiffusionGemma utilize parallel denoising techniques instead of the traditional word-by-word generation method. This allows them to produce responses more efficiently. The shift to a diffusion model approach means that the AI fills a text block with placeholder tokens and iteratively refines them into coherent sentences, similar to how image generators like Stable Diffusion work.

Despite Mercury 2's speed and higher test scores, it is important to note that it operates as a closed-weight model available through a paid API, while DiffusionGemma is free and open-source, located on Hugging Face. This could influence the adoption of each model depending on user needs and budgets.

Inception Labs maintains that Mercury 2 leads in terms of quality, performance, and cost-effectiveness, positioning it at the front of the current generation of AI tools. The company states:

Looking ahead, developers will be keen to assess how Mercury 2 performs in real-world applications beyond benchmarks, especially given that it is already being integrated by companies such as Augment Code, which replaced Anthropic’s Claude Opus 4.7 with Mercury 2 for enhanced AI coding tasks.