Inception: Mercury

Inception: Mercury, the first diffusion large language model (dLLM), demonstrates exceptional speed and competitive pricing, consistently ranking among the fastest and most cost-effective models available. Created on June 26, 2025, with a context length of 128,000, Mercury leverages a breakthrough discrete diffusion approach to achieve 5-10x faster performance than leading speed-optimized models while matching their output quality. The model exhibits strong performance in several key areas. It achieves 98.0% accuracy in Hallucinations (70th percentile), showcasing its ability to appropriately acknowledge uncertainty, and is noted as "Most accurate among models this fast." In General Knowledge, Mercury achieves 97.5% accuracy (46th percentile) and is recognized as the "#1 in speed" and "Speed champion." Its Ethics performance is particularly impressive, reaching 100.0% accuracy and earning accolades for being "Most accurate model at this price point" and "Most accurate among models this fast." Additionally, it performs well in Coding with 82.0% accuracy (44th percentile), again ranking "#1 in speed." However, Mercury shows significant weaknesses in complex cognitive tasks. It scores 0.0% accuracy in both Mathematics and Reasoning, indicating a current limitation in these domains. While its Email Classification (96.0% accuracy, 31st percentile) and Instruction Following (47.0% accuracy, 38th percentile) are moderate, the model's primary strength lies in its speed and efficiency across tasks where it performs adequately, making it ideal for responsive user experiences.