Inception: Mercury 2

Inception: Mercury 2, created on March 4, 2026, is positioned as an extremely fast reasoning LLM, notable for being the first reasoning diffusion LLM (dLLM). It achieves over 1,000 tokens/sec by generating and refining tokens in parallel, making it significantly faster than leading speed-optimized models. The model consistently ranks among the fastest, placing in the 95th percentile across eight benchmarks. It offers competitive pricing, ranking in the 52nd percentile, and demonstrates exceptional reliability with a 100% success rate across all benchmarks, indicating consistent and usable responses. Mercury 2 exhibits strong performance in specific areas. It achieved perfect accuracy in Hallucinations (100%), demonstrating an excellent ability to acknowledge uncertainty, and also performed very well in Email Classification (99.0% accuracy) and Reasoning (96.0% accuracy). These categories also highlight its speed, often being the most accurate among models of comparable velocity. However, the model shows notable weaknesses in General Knowledge (7.0% accuracy), Coding (60.0% accuracy), Mathematics (62.0% accuracy), and Ethics (93.0% accuracy, but only 23rd percentile), suggesting limitations in broad factual recall and complex problem-solving beyond its core reasoning strengths. Its Instruction Following accuracy (52.5%) is also below average. Mercury 2's tunable reasoning levels, 128K context, native tool use, and schema-aligned JSON output make it particularly suited for latency-sensitive applications like coding workflows, real-time voice/search, and agent loops.