Mercury-2

State-of-the-art diffusion based reasoning model.

Mercury 2 is the fastest reasoning LLM and the first reasoning dLLM — delivering 5–10x faster inference than speed-optimized autoregressive models like Claude 4.5 Haiku and GPT-5 Mini. Built on the diffusion LLM architecture Inception pioneered, Mercury 2 generates output through parallel refinement rather than sequential decoding, achieving ~1,000 tokens per second on NVIDIA GPUs. It's production-grade, API-accessible, and priced at a fraction of comparable models.

Mercury 2 is OpenAI client API compatible and supports a 128K context length, tool calling, structured outputs - plus a rich set of tuning options to tailor deployments to your needs.