Artificial Analysis Video Leaderboards Now Label Derived Models and Allow Hiding Them

Artificial Analysis announced that its video leaderboard will now label derived models and allow users to hide them. "Derived models" here means models post-trained on top of another model. The organization says this type of model is increasingly common on its leaderboard.

This update is noteworthy not because it adds a filter button, but because it acknowledges a structural blind spot in the leaderboard: when multiple ranked models share the same base model, the "number of independent models" reflected by the rankings may be overstated. Artificial Analysis's own example is that in the top five of the AA-Video-T2V v2.0 leaderboard, two models—Utopai X and MiniMax H3 Max—are built on MiniMax H3—that is, two of the top five are same-origin.

A natural question is: since they originate from the same base model, why not just remove them from the leaderboard? Artificial Analysis's answer is that derived models remain on the leaderboard by default because they are different models from their base models.

The organization further explains that derived models may be clearly better or worse than their base models in specific capabilities and use cases, and often differ from their base models in pricing, speed, and availability. It argues that these differences matter to anyone choosing a model. In other words, keeping them by default is a trade-off that prioritizes "information completeness": rather than deciding for users what counts as an independent model, it marks the same-origin relationship and lets users judge for themselves.

This update lands as two visible interaction changes. First, derived models now display their base model's logo next to their name, presenting the "lineage" directly within the leaderboard row. Second, users can choose "hide derived models" to rank only base models.

These two actions correspond to two different reading needs: those who want to see the full ecosystem view look at the complete leaderboard with logos; those who want to compare base models themselves use the toggle to filter out derived entries. It should be noted that the source does not provide a specific example of how the leaderboard rankings change after "hiding derived models," nor does it state whether this feature affects historical leaderboards or exported data.

The announcement post shows 7,245 views, and the follow-up questions in the comments point precisely to two weak spots in the labeling mechanism. Commenter Kepler 22b asks: if a lab does not disclose it, how can one tell that a model was post-trained on MiniMax H3? Commenter NotaDEV asks whether the pricing difference between derived models and MiniMax H3 itself is very large.

Neither question is answered in the source. The source does not state the specific criteria or detection method for derived models, nor does it give specific figures for derived models versus base models in pricing, speed, and availability. From this one can infer that identifying derived models may rely on lab disclosure or third-party judgment, but how exactly Artificial Analysis determines a derived relationship is not explained by the source.

Commenter Rohan raises a sharper observation: derived models inherit the blind spots of their base models, so the labeling also shows how many independent entries the top five actually contains.

This claim elevates the labeling from a "classification tool" to a "measure of leaderboard credibility." If two of the top five are based on MiniMax H3, then those five seats represent fewer than five independent technical paths. It must be emphasized that this is the commenter's view, not an official statement from Artificial Analysis; the organization itself only states that derived models differ from their base models and that the differences matter to those choosing models, and does not make any judgment about the "number of independent entries."

Separating the confirmed from the unconfirmed makes the weight of this update clearer. What is confirmed: the leaderboard now labels derived models and allows hiding them; derived models display their base model's logo next to their name; users can choose to rank only base models; the announcement post shows 7,245 views.

What belongs to the organization's own claims and awaits external verification: the number of derived models is "growing"; in the top five, Utopai X and MiniMax H3 Max are based on MiniMax H3; derived models may be clearly better or worse than base models in specific capabilities; they often differ in pricing, speed, and availability. The source is only an announcement post and its comments, with no complete leaderboard or methodology document, and no complete ranking or scoring data for AA-Video-T2V v2.0.

There are also several dimensions the source does not cover at all: whether the growth trend in the number of derived models is supported by data (the source only says "growing"); whether the labeling feature covers Artificial Analysis's other leaderboards; and information about the developer or publisher of MiniMax H3. These gaps mean that what can currently be confirmed is that "the labeling mechanism has launched," not that "the derived-model ecosystem has been quantified."