Mercury 2: The fastest reasoning LLM, powered by diffusion (www.inceptionlabs.ai)

Today, Inception announced Mercury 2, the fastest reasoning language model to date, designed to enhance the responsiveness of production AI applications. Unlike traditional models that rely on autoregressive, sequential decoding—processing one token at a time—Mercury 2 employs a novel diffusion-based approach, enabling it to generate multiple tokens simultaneously. This leap in technology results in generation speeds exceeding 1,009 tokens per second on NVIDIA Blackwell GPUs, achieving over five times the speed of current leading models while maintaining competitive quality. The significance of Mercury 2 lies in its capability to redefine performance for latency-sensitive tasks such as coding, interactive dialogue, and intelligent search systems. By dramatically reducing latency in agentic workflows and real-time interactions, it enhances user experience, making AI suggestions feel more intuitive. With features like tunable reasoning, 128K context, and native tool use, Mercury 2 not only meets but exceeds the demands of modern AI applications—allowing developers to seamlessly integrate advanced reasoning capabilities into their existing workflows without extensive modifications. This innovation sets a new benchmark in the AI/ML landscape, particularly for businesses focused on real-time engagement and performance optimization.