Yi-Lightning

Yi-Lightning is a speed-optimized language model developed by 01.AI, the Chinese AI company founded by Kai-Fu Lee. Built on an enhanced Mixture-of-Experts (MoE) architecture with advanced expert segmentation and optimized KV-caching techniques, it achieves over 200 tokens per second on consumer GPUs (RTX 4090) and 500+ tokens per second on H100s.

Upon its debut in October 2024, Yi-Lightning achieved 6th place overall on the Chatbot Arena leaderboard based on real-world human evaluation, with top rankings in specialized categories: 2nd to 4th place in Chinese language, mathematics, coding, and hard prompts. This placed it competitively against models from OpenAI, Anthropic, and Google.

Trained at a cost of just $3 million using 2,000 H100 GPUs, Yi-Lightning is 70-80% more cost-effective than US frontier models for coding and mathematical workloads. This extreme cost efficiency, combined with its competitive performance, exemplifies the broader trend of Chinese AI labs achieving frontier-level results at dramatically lower training budgets.