Mercury 3: Diffusion Model Outperforms Frontier LLMs
Mercury 3 is a diffusion-based language model achieving performance parity with GPT-6 Luna and Claude Haiku 5.5 while generating at 1,068 tokens per second on NVIDIA GPUs. The model features 1M-token context, up from 260K in Mercury 2.5, and costs $0.20 per million input tokens and $0.70 per million output tokens, with launch pricing at 50% discount. Intelligence benchmark scores improved by three times from Mercury 1 to Mercury 3.
Early adopters deploy Mercury 3 for latency-sensitive applications. Web navigation agents solve 50% of Mind2Web tasks, outperforming GPT-6.1-Sol. Tutoring systems provide instant feedback without compromising reasoning. A smart glasses search implementation achieved 0.9-second median generation time and 2.2-second end-to-end response. Customer support agents maintain accuracy while reducing latency. The results demonstrate that diffusion models follow scaling laws comparable to autoregressive approaches.
Mercury 3 is part of a product family including Mercury Voice for voice applications and Mercury Decide for high-volume classification. The model is available via Inception API Platform, OpenRouter, Baseten, and Vercel with OpenAI API compatibility.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.