We need your help to keep this website free. You can 
A November 7 update to X’s public recommendation code gives its video carousel a named click prediction model. The change appeared in the repository maintained by xAI. It identifies a model the system can reference, but the public code alone does not establish that the carousel or that model is running for users.
Source: Berke Citak / Unsplash, Unsplash License. Cropped archive photograph of the X app icon on a phone home screen.
The commit replaced an empty value in the carousel model field with the name of a specific model. That setting lives alongside controls for enabling the carousel and its click model. Both switches still default to false in the published file. A separate production configuration could give them other values, so those defaults are not a view of what an account currently sees.
The new name is one setting in a larger system
The model field sits in Home Mixer, the part of the recommendation code that assembles the feed. A named model can be selected if the system uses that feature. The name itself contains no measurements and gives no explanation of which visual characteristics might influence the score. The same change added a five-minute cache setting for ad history. That is separate from the carousel model field and does not tell a video publisher how placements are selected.
The last X code update that added mention data illustrated a similar limit. A repository can show that an input or setting exists without showing how broadly it is used. In this case the release does not specify the audience, the dimensions of the video cards, or whether the prediction applies to a click on the video itself or another element of a card.
A click and a view answer different questions
Click-through rate compares clicks with impressions. A video can collect a click because its cover makes a promise and then lose the viewer when the opening seconds fail to deliver it. Conversely, someone who encounters an autoplaying clip can watch without tapping the cover. An observed change in watch time would not by itself reveal whether a click model or the thumbnail caused it.
If a carousel placement becomes visible, its card crop can be checked against the cover already used on the regular feed. Look at whether the subject and text still fit before creating a separate export. For measurement, compare covers on the same type of placement and keep video length and the opening shot consistent. The X API exhibit is a separate place to look up what an integration exposes. It cannot establish that a video appeared in a particular user’s feed.
For now, the repository provides a timestamped change and a model identifier. A creator would need a visible carousel and a placement-specific denominator before treating the change as evidence about thumbnail performance.
