Yi-Lightning - The latest high-performance, high-speed flagship model from 01Wuwu.

What is Yi-Lightning?

Yi-Lightning, the latest flagship model released by Zero-One-Way Technology Co., Ltd., has achieved remarkable results on the internationally authoritative blind benchmarking list LMSYS, surpassing OpenAI's GPT-4o-2024-05-13 and Anthropic Claude 3.5 Sonnet, ranking sixth in the world and first in China. This achievement marks the first time that a Chinese large-scale model has surpassed OpenAI's GPT-4o in the global arena, demonstrating China's strength and progress in the field of artificial intelligence.

The Yi-Lightning model also demonstrated outstanding performance on multiple leaderboards. In the Chinese leaderboard, it surpassed other high-quality models from both domestic and international sources, tying for second place globally with models such as o1-mini. In the multi-turn dialogue leaderboard, Yi-Lightning ranked third, and in the mathematics and code leaderboards, it achieved third and fourth place, respectively.

Yi-Lightning has also achieved breakthroughs in inference speed and cost. Compared to its predecessor, the flagship model Yi-Large, Yi-Lightning's peak generation speed has increased by nearly 40%, and the first-pack time has been halved. The inference cost of Yi-Lightning has been further reduced, costing only 0.99 yuan per million tokens, approaching the lowest price in the industry.

Main functions of Yi-Lightning

• Inference speed and costYi-Lightning offers a significant improvement in inference speed compared to its predecessor, the flagship model Yi-Large, with a peak generation speed increase of nearly 40%. Inference costs have also been further reduced, costing only 0.99 yuan per million tokens.
• AI 2.0 Digital Human Solution0150 has launched an AI 2.0 digital human solution based on the Yi-Lightning model, focusing on scenarios such as retail and e-commerce. This solution includes large-scale character models, live-streaming voice models, and e-commerce script models, possessing capabilities such as motion training, facial expression generation, multilingual and emotional expression, and intelligent dialogue. In practical applications, a travel and hospitality company saw its GMV increase by 170% after implementing the solution.
• Industry-wide solutionsThe Yi-Lightning model is also applied to industry-wide solutions for everything, including the basic model, which are paired with practical tools such as RAG and Function Calling. It has already been implemented in retail, healthcare, education, logistics, and other fields, covering application scenarios such as AI search, AI productivity tools, and AI intelligent inspection.
• Model architecture innovationYi-Lightning adopts the Mixture of Experts (MoE) hybrid expert model architecture, which introduces a hybrid attention mechanism and a dynamic Top-P routing mechanism during model training. It innovatively provides a standardized base model with a higher starting point, enabling faster, more efficient, and lower-cost training of customized models.
• Speed ReasoningYi-Lightning has a very fast inference speed. Based on the dynamic Top-P routing mechanism, it can dynamically and automatically select the most suitable combination of expert networks according to the difficulty of the task, balancing inference cost and model performance.
• Multi-stage trainingYi-Lightning employs a multi-stage training model, emphasizing data diversity in the early stages and focusing on richer, more informative data in the later stages. This training method helps the model absorb knowledge from different stages, and the training speed and stability are ensured by adjusting the batch size and learning rate (LR).

The technical principle of Yi-Lightning

• MoE Hybrid Expert Model ArchitectureYi-Lightning employs a Mixture of Experts (MoE) architecture. This architecture combines multiple expert networks to handle different tasks, allowing the model to dynamically select which expert networks to activate based on the task's difficulty, balancing inference cost and model performance. During training, all expert networks are activated; during inference, the model selectively activates the more suitable expert networks.
• Hybrid Attention MechanismYi-Lightning optimizes the hybrid attention mechanism by replacing the traditional full attention with sliding window attention only in some layers of the model, reducing computational costs while maintaining efficient processing capabilities for long sequence data.
• Cross-Layer Attention (CLA)Yi-Lightning introduces a cross-layer attention mechanism, which allows models to share key and value headers across different layers, reducing the need for storage resources and improving the model's inference efficiency.
• Dynamic Top-P RoutingYi-Lightning dynamically and automatically selects the most suitable combination of expert networks based on the difficulty of the task, without the need for manual intervention. This enables the model to more intelligently adapt to various task requirements and achieve ultra-fast reasoning.

Yi-Lightning's project address

• Project official websiteplatform.lingyiwanwu.com

Application scenarios of Yi-Lightning

• Translation scenariosYi-Lightning excels in the translation field, handling language understanding and generation, cross-linguistic capabilities, and context awareness to provide high-quality translation services. In comparisons with multiple models, Yi-Lightning's translation capabilities are clearly demonstrated, exhibiting precise word choice and literary flair.
• Retail e-commerce live streaming scenariosLingyiwu's AI 2.0 digital human solution focuses on scenarios such as retail and e-commerce. Based on the technology provided by Yi-Lightning, it enables functions such as interactive bullet comments, product information extraction, and real-time script generation. After integrating with Yi-Lightning, the digital human's real-time interactive effect is better, the scripts are smoother, and the responses are more accurate.
• Enterprise-level solutionsYi-Lightning is also applied to enterprise-level solutions under the To B strategy of "zero-one-things" to provide enterprises with customized AI services and help them improve efficiency and revenue.
• Multilingual processingOn the Chinese rankings, Yi-Lightning demonstrated powerful Chinese processing capabilities, which are comparable to those of top international models.
• Mathematics and code generationYi-Lightning achieved third and fourth place in the math and coding categories, respectively, demonstrating its strong capabilities in these areas.
• Long and difficult questionsYi-Lightning also excelled in handling long and difficult questions, achieving an outstanding fourth place in the world in both categories, demonstrating its ability to solve complex problems.