Kuaishou open-sources the Kolors text-to-image model

Background

A text-to-image model takes a written description and produces a picture that fits it. Building one requires large collections of images paired with descriptions, and the computing power to train on them. Because such models are costly to build, the question of how they are released matters a great deal. Some are offered only through the company's own interface; others are published as open source, meaning the code, and often the trained weights, can be downloaded and run by anyone. An open release lets outside researchers inspect how a model behaves, adapt it to their own data and build on top of it, rather than only sending queries to a service someone else controls.

What happened

On 6 July 2024, the Chinese short-video company Kuaishou released Kolors, its text-to-image model ^[1]. It was not a standalone announcement: the model was presented at a forum held in Beijing during WAIC 2024, under the title "New AI • New Applications • New Ecosystem". Kolors appeared there alongside two other Kuaishou models, KwaiYii and Kling ^[2].

Key details

The release separated the model's openness into two parts. The code was published under Apache-2.0, a permissive license that allows others to use, modify and redistribute the software as long as the license terms are kept. The trained weights were opened for research ^[1].

For research use, the weights are the more consequential half. Weights are the numbers that make up a trained model; with them in hand, a lab can run the model on its own machines, test it on its own prompts and fine-tune it, instead of only calling an outside service. Publishing code under a permissive license lowers the barrier further, since others can build their own tools around the model without negotiating a separate agreement.

Kuaishou also made a performance claim: it said Kolors took second place globally in a text-to-image evaluation run by BAAI ^[2]. Rankings of this kind compare models on a fixed set of prompts and scoring criteria. They are one of the few public signals available for judging image models against one another, though they measure only what the benchmark asks, and results can shift as new models are added.

In its announcement, Kuaishou described the release as having been "officially open-sourced to foster industry vitality" ^[2] — framing openness itself, rather than only the model's output quality, as the point of the release.

Why it matters

For outside researchers, a model released with downloadable weights is something to be examined and reused, not just a product to be tried. The claim of a second-place ranking gave that release a reference point: readers could compare Kolors against other models on a public evaluation rather than take a vendor's word for its quality ^[2].

Kolors also arrived as part of a larger showing. Appearing at the same Beijing forum as KwaiYii and Kling meant a company known mainly for short video was presenting several models across different areas of AI at once ^[2]. For readers following open model releases, that combination — an openly licensed model, a stated benchmark position and a coordinated launch — is a useful example of how such releases are typically packaged.