奈亨智能
Pursuing Mercury-level efficiency in AI reasoning — achieving autoregressive reasoning with diffusion-speed generation.
We're building an open-source implementation of breakthrough AI architectures that combine the best of autoregressive and diffusion models.
Mercury Coder 2 demonstrated that autoregressive models can achieve diffusion-level speeds while maintaining coherent reasoning. We believe this architecture should be accessible to everyone.
Faster inference than traditional autoregressive models
Research and implementation for the community
Documented approach for others to follow
Studying Mercury's parallel decoding mechanism and understanding how it achieves diffusion-style generation while preserving autoregressive reasoning capabilities.
Building our version from first principles, focusing on the key innovations: parallel token prediction, efficient attention patterns, and coherent multi-step reasoning.
Sharing our findings, code, and models with the community so others can build upon efficient reasoning architectures.