Inception: Mercury 2.5 Preview

Inception: Mercury 2.5 Preview, created on August 31, 2026, is positioned as a groundbreaking diffusion LLM (dLLM) that generates and refines tokens in parallel, aiming for unparalleled speed. This model consistently ranks among the fastest and most competitively priced models across eight benchmarks, achieving an "Infinityth percentile" in both speed and price. Its reliability is exceptional, boasting a 100% success rate across all benchmarks, indicating minimal technical failures. Performance across benchmarks reveals a model with significant strengths. Mercury 2.5 achieved perfect 100.0% accuracy in General Knowledge, notably being the most accurate model at its price point and among models of comparable speed. It also demonstrated strong performance in Coding (95.7% accuracy, 91st percentile), Reasoning (98.0% accuracy, 91st percentile), and Mathematics (97.0% accuracy, 97th percentile), showcasing robust analytical and problem-solving capabilities. A notable weakness is its 0.0% accuracy in Instruction Following, suggesting a critical area for improvement. Hallucinations were relatively low at 98.0% accuracy, indicating a good ability to acknowledge uncertainty.