01.AI open-sources Yi-1.5 under Apache 2.0

Background

The original Yi models had established 01.AI as a publisher of open bilingual language models. In spring 2024 the company was approaching the first anniversary of its founding in Beijing, and it used that occasion to announce several models. Kai-Fu Lee, who founded the company and serves as its chief executive, was part of the announcement ^[3].

An open release is a decision about access as well as technology. When a company publishes weights under an open licence, anyone can download them and build on them. The company's choice of licence, and the sizes it releases, decide who can use the models and for what.

What happened

On 13 May 2024 01.AI open-sourced the Yi-1.5 series in three sizes: 34B, 9B and 6B ^[1]. The series has base models and chat models for each size ^[3]. The release was presented as part of the company's first-anniversary event, which also launched the closed Yi-Large model ^[3].

Yi-1.5 was continuously pretrained on the original Yi models with 500 billion more tokens and then fine-tuned on 3 million samples ^[2]. These two stages are the core of the upgrade. The code and weights were released under the Apache 2.0 licence ^[1].

The company also said Yi-1.5-9B-Chat led similarly sized open models ^[2]. Readers should treat this as a claim by the company rather than an independent result.

Key details

The models come in base and chat versions, so developers can choose between a model for continuing text and a model that has been tuned for conversation. Variants with 16K and 32K context windows were also released, such as Yi-1.5-34B-32K ^[2]. These windows describe how much text the models can read at once.

The Apache 2.0 licence is the most important legal detail. It allows use, modification and redistribution, including for commercial purposes, provided that the licence notices are kept. Because the original Yi-34B had been released for research and non-commercial use, this licence change is one of the main differences between the two releases.

Impact

The Yi-1.5 release widened access to the Yi family. Developers who could not use the original models in commercial products could now build with the new ones, and researchers could download the weights, test them and publish their own results. The release also showed the company's approach of keeping an open line of models alongside its closed flagship.

For readers new to the field, the event is a useful case study. It shows how a version number, a set of sizes and a licence together describe what a model is and who can use it. It also shows that companies often publish their own benchmark results at launch, and that these results are most useful when they are checked against independent tests.

The release also belongs to a longer timeline. Later in 2024 the company released Yi-Coder and Yi-Lightning. Reading the May launch together with those later releases gives a fuller picture of how the company's line of models developed.