Inception's Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5 at a fraction of the wait
decoder ·
AI startup Inception has released Mercury 3, a diffusion-based language model that claims to outperform competitors like GPT-6 Luna and Claude Haiku 5.5 in speed and specific benchmarks. Unlike traditional autoregressive models that generate text word-by-word, Mercury 3 refines a rough draft of the entire response simultaneously, allowing it to achieve over 1,000 tokens per second on Nvidia GPUs. Internal testing indicates that Mercury 3 leads in eight out of twelve tests, showing significant advantages in instruction following and autonomous coding tasks. However, rivals still hold strengths in longer coding challenges and information retrieval within extensive texts. The model is designed for real-time applications such as AI agents, smart glasses, and customer support, where low latency is critical. Early partners report substantial improvements, with search responses on smart glasses dropping from five seconds to under one second. While launch pricing is currently discounted, standard rates may eventually exceed those of some competitors for shorter inputs. Inception, backed by major tech investors, positions this architecture as a scalable alternative to standard language models.
decoder