Yi-Large

What it is

Yi-Large is a closed-source large language model with a parameter scale of about one hundred billion (千亿), launched by 01.AI on 13 May 2024 at the company's first-anniversary event^[1]. Developers reach it through 01.AI's API platform, which opened globally on the same day^[1].

It did not arrive alone. The platform also carries Yi-Large-Turbo, Yi-Medium, Yi-Medium-200K, Yi-Vision and Yi-Spark, a set of models that differ in size and in what they handle^[1]. Calling Yi-Large "closed" means the weights are not published for others to download and run; access is sold as a service instead. That is the opposite arrangement from an open-weights release, where anyone can take the model and host it themselves.

How it was described

QbitAI reported that Yi-Large was presented as a "hundred-billion-parameter SOTA base model" (千亿参数SOTA基座模型) offered on the Yi API platform^[2]. A base model is a general-purpose model trained on broad data rather than on one narrow task; it is the starting point that applications and further tuning build on. The description therefore places Yi-Large as the foundation of 01.AI's model lineup rather than as a single finished product.

Benchmarks

The numbers 01.AI emphasised were mostly from public leaderboards. On Stanford's AlpacaEval 2.0 leaderboard, Yi-Large ranked second in the world on length-controlled win rate, behind GPT-4 Turbo, and first on raw win rate^[1][3]. Those two figures measure slightly different things: the length-controlled version corrects for the tendency of evaluators to favour longer answers, while the raw win rate does not. On the April 2024 SuperCLUE benchmark, it ranked first among Chinese models^[1].

In 01.AI's own evaluation, the company said Yi-Large beat GPT-4, Claude 3 and Gemini 1.5 on most metrics, ranking first on HumanEval and MATH^[1][2]. HumanEval tests whether a model can write working code and MATH tests mathematical problem solving, so those two results are aimed at reasoning and programming rather than at open conversation. Separately, instruction-following results on AlpacaEval, MT-Bench and Tsinghua's AlignBench were reported as better than the world's top five models^[3].

These comparisons come from the company's own testing and from leaderboards it chose to cite, so they should be read as claims made at launch rather than as an independent audit of the model.

Price and access

At launch, the API price was 20 yuan per million tokens, which 01.AI described as about one third of GPT-4 Turbo's price^[2]. Tokens are the units of text that models read and write, so per-token pricing is how API usage is normally billed: the more text a developer sends and receives, the more tokens are consumed. Pricing at roughly a third of a leading competitor's rate was part of how the model was marketed to developers choosing which API to build on^[2].

Released alongside other products

Yi-Large came out together with the open Yi-1.5 models and with Wanzhi, a consumer productivity app^[1]. The pairing is notable in how the lineup was split: open models that anyone can download, a closed flagship sold through the API, and a finished application for end users^[1].

01.AI also announced at the same event that it had begun training a next-generation Yi-XLarge MoE model aimed at GPT-5-level performance, and said that in early evaluations this model traded wins with Claude-3-Opus and GPT4-0409^[3]. MoE, or mixture of experts, is an architecture in which only part of the network is used for each input, which can make a very large model cheaper to run than its total parameter count would suggest. That model was described as in training, not as released.

Where it sits

Within 01.AI's own lineup, Yi-Large was the closed flagship at the top of a family that also included open Yi-1.5 models and smaller or specialised API models^[1]. Its launch package combined three things a model needs to be used in practice: benchmark results on widely followed leaderboards, a published price per token, and an API platform open internationally from day one^[1][2][3]. The later Yi-XLarge MoE announcement indicates that Yi-Large was treated as a step in an ongoing series rather than an endpoint^[3].