Mercury 2

Mercury 2 redefines language generation by using parallel refinement instead of sequential decoding. This innovative approach allows for simultaneous token generation, achieving speeds over 1,000 tokens per second. This high velocity ensures reasoning-grade quality can be delivered within strict latency budgets, essential for complex operational loops like agents, retrieval pipelines, and extraction jobs. Its unique diffusion-based reasoning fundamentally alters the speed curve, providing significantly faster responses without compromising on analytical depth. Designed for production environments, Mercury 2 offers competitive quality alongside its unparalleled speed. Key features include tunable reasoning capabilities, expansive 128K context window, native tool utilization, and schema-aligned JSON output. Targeted at engineering and product teams building high-performance applications, Mercury 2 unlocks real-time interactivity and responsiveness, even under high concurrency. It's ideal for tasks where quick, accurate language processing is critical, enabling instant and consistent turn-to-turn behavior for critical operational systems.