Inception Labs Introduces Mercury 3: High-Speed Diffusion LLM
Inception Labs has introduced Mercury 3, a large-scale diffusion model that achieves generation speeds of over 1,000 tokens per second. The company claims it outperforms select competing frontier models on key benchmarks. This release highlights the ongoing advancement of non-autoregressive diffusion architectures in generative AI, offering significantly faster throughput and lower latency than traditional models. Such improvements could drastically reduce serving costs and enhance real-time application responsiveness. Mercury 3 operates as a diffusion-based large language model capable of running 4 to 8 times faster than traditional alternatives while maintaining competitive quality. Users have noted its strong multi-turn interaction capabilities and effective code and interface generation performance.
## BACKGROUND
Traditional large language models are autoregressive, meaning they generate text one token at a time in sequence. In contrast, diffusion models generate data iteratively through a denoising process, potentially allowing large chunks or the entire sequence to be synthesized in parallel for much higher speeds.