Mercury 2 tops Artificial Analysis speed rankings

Mercury 2's parallel diffusion architecture represents a fundamental paradigm shift that challenges the sequential generation bottleneck of traditional autoregressive models, proving that extreme speed and reasoning are not mutually exclusive.

• –**Parallel Refinement:** Adapting diffusion techniques to text generation enables parallel token refinement, drastically increasing throughput.
• –**Unprecedented Throughput:** Exceeding 1,000 tokens per second on NVIDIA Blackwell GPUs makes it a massive game-changer for latency-sensitive agentic workflows.
• –**Architectural Trade-offs:** While highly competitive on coding and structured JSON generation, its parallel generation architecture must continue to prove its robustness on multi-step complex reasoning compared to state-of-the-art autoregressive transformers.