Inception Labs unveils Mercury 3, a 1,000-token-per-second diffusion LLM

Summary

Inception Labs introduced Mercury 3, its most capable diffusion-based language model to date. By the company's own benchmarks, it matches or beats GPT-6 Luna and Claude Haiku 5.5 in quality while generating 1,068 tokens per second on NVIDIA GPUs. Inception says the model is more than 40x larger than Mercury 1 yet just as fast, arguing that diffusion models follow the same scaling laws as autoregressive ones. Target workloads are latency-sensitive agent tasks such as web navigation agents, real-time search, and voice tutoring. All quality figures are the company's self-evaluations.

Why it matters for our work

Agents wait on the model at every step, so generation speed directly sets task completion speed. More cheap, ultra-fast model options let agents try more approaches within the same time and budget, speeding up human-AI collaboration.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.