Inception just dropped a new large language model that takes a fundamentally different approach to generating text. Mercury 2.5, the company’s latest diffusion-based LLM, delivers a 40% intelligence improvement over its predecessor while maintaining the kind of speed and cost profile that makes enterprise CFOs smile.
The model processes 1,107 tokens per second on standard NVIDIA GPUs.
What makes diffusion models different
Most large language models you’ve interacted with, think GPT or Claude, generate text one token at a time in sequence. They’re autoregressive, meaning each word depends on the one before it. Diffusion models work differently. They generate multiple tokens simultaneously, more like how an image diffusion model creates a picture by gradually refining noise into something coherent.
Mercury 2.5 comes with a 260K token context window, which means it can process roughly the equivalent of a 500-page book in a single prompt. It also supports tunable reasoning levels, letting developers dial the model’s thinking depth up or down depending on whether they need deep analysis or quick responses. Parallel tool calls and structured JSON output round out the feature set.
