Inception Labs has introduced Mercury 2, described as the world’s fastest reasoning language model; Mercury 2 diffusion LLM speed and benchmarks are presented by the company as a central claim. Benchmarks state Mercury 2 generates about 1,000 tokens per second, compared with Claude Haiku 4.5 Reasoning at 89 tokens per second and GPT-5 Mini at 71 tokens per second. Those figures are provided as key speed benchmarks for comparing reasoning-focused models.
Mercury 2’s benchmark results in the provided material list a 90% score in AIME 2026 and a 77% score on GPQA, presenting those two test outcomes as the primary performance figures for the model.
In the AIME 2026 rundown included in the material, DiffusionGemma is listed at 69.1% while Gemma 4 is listed at 88.3%, and those three models are presented together for that specific benchmark.
For GPQA, the provided figures show Mercury 2 at 77% and DiffusionGemma at 73.2%, with GPQA called out separately from AIME 2026.
The material also states that Google’s developer guide recommends standard Gemma 4 for applications that demand maximum quality and that DiffusionGemma trails Gemma 4 across the board.
The provided context additionally names Claude Haiku 4.5 Reasoning and GPT-5 Mini among the products mentioned alongside these benchmark results.
Provided material states that for non-technical users diffusion models provide instant autocomplete, rapid iterations on code or plans, and sub-agents that can handle the boring high-volume work. The material presents these features as practical, user-facing benefits for tasks such as editing and planning without explaining underlying mechanisms. The provided context mentions these user benefits alongside Mercury 2 as an example of a diffusion model.
The Augment Code case study in the provided material reports that Mercury 2 substituted for Claude Opus 4.7, resulting in an 82% drop in latency and a 90% cut in cost while maintaining the same output quality. The provided material also states that Mercury 2 was built on research from Stefano Ermon, a Stanford professor. The provided content reports Inception Labs raised $50 million in funding that drew backing from Nvidia’s venture arm and individual investors Andrew Ng and Andrej Karpathy.
The section compiles the practical benefits, a customer case study, and the model’s research and funding background as reported in the provided material. Technical mechanisms and direct quotes are not included in this summary.
The provided material presents Mercury 2 as a fast reasoning language model introduced by Inception Labs and situates it within the diffusion-era AI landscape. The material links Mercury 2 to diffusion-model user benefits such as instant autocomplete, rapid iterations on code or plans, and sub-agents for high‑volume work, and it records that the model builds on academic research and attracted industry funding.
