Yi-Lightning

What it is

Yi-Lightning is the flagship model that 01.AI launched in October 2024 ^[1][3]. It is a large language model built on an enhanced Mixture-of-Experts (MoE) architecture. A Mixture-of-Experts model splits its network into many specialised parts and activates only some of them for each input. The company's version uses fine-grained expert segmentation and routing, meaning the experts are divided into smaller units and a routing step decides which ones handle each input. It also uses optimised KV-caching, a technique that stores intermediate results from earlier tokens so the model does not need to recompute them for every new word ^[1].

Yi-Lightning succeeded Yi-Large as 01.AI's top API model ^[3].

How it was trained

According to the technical report, Yi-Lightning was trained in several stages. Pre-training teaches the model from a large body of text. Supervised fine-tuning then teaches it to respond to instructions with examples of good answers. Reinforcement learning from human feedback (RLHF) further adjusts the model based on human judgements of its responses, using a reward model that scores answers ^[1]. The report also describes multi-stage training, synthetic data construction and reward modelling as parts of the process ^[1].

Synthetic data means examples generated by models rather than written by people, which can be used to expand training material on topics that are hard to collect at scale.

Results

The model ranked sixth overall on the LMSYS Chatbot Arena, and placed between second and fourth in the Chinese, Math, Coding and Hard Prompts categories ^[1]. The Chatbot Arena is a public platform where people vote on answers from two anonymous models in a blind test, and the votes are turned into a ranking. Each category groups questions by theme, and Hard Prompts groups more demanding questions.

Kai-Fu Lee said that Yi-Lightning placed sixth in the world and first among Chinese models on LMSYS's blind-test ranking, ahead of GPT-4o-2024-05-13 and Claude 3.5 Sonnet ^[2]. The model names here are those of the competing systems at the time of the ranking.

In the UC Berkeley SkyLab and LMSYS ranking of model makers, 01.AI tied for third with xAI, whose model is Grok-2. OpenAI and Google ranked ahead of them ^[3]. This ranking looks at companies rather than individual models.

Price

The API price at launch was 0.99 yuan per million tokens. Lee said the price included a small margin above cost, rather than being set below cost ^[2]. A token is a small unit of text, often a word or part of one, and API prices are usually quoted per million tokens.

Wikipedia gives a different comparison for the inference price, which is the cost of running the model to generate answers: 14 US cents per million tokens, against 26 cents for OpenAI's o1-mini ^[3]. The two figures come from different sources and measure the price in different currencies, so readers should check the source of each.

Significance

Yi-Lightning showed that a model from a Chinese company could rank at the top of a widely watched public arena while also being cheap to use. Public rankings were the main external check on how the model compared with others.

The model also marked a turning point in the company's history. In March 2025 01.AI stopped pre-training large language models, according to Wikipedia, and its later work moved toward tailored business solutions ^[3].

Safety

The technical report describes RAISE, the Responsible AI Safety Engine, a four-part safety framework covering pre-training, post-training and serving the model ^[1]. Pre-training is the stage when the model learns from text, post-training is the stage when it is adjusted for use, and serving is the stage when it answers users. A framework that covers all three stages treats safety as part of the whole life of the model rather than a single check at the end.

Publication

The technical report was posted to arXiv on 2 December 2024 ^[1]. arXiv is an open archive of scientific papers, and posting a report there allows researchers to read and cite the details.