New Meta Artificial Intelligence models, including Scout light and medium-format Maverick, which are exposed to GPT-4O and Gemini 2.0 Flash in popular tests, have become the object of scandal through testing with test results. As it turned out, the advertising of the public version of the model was used statistics of the experimental version, inaccessible to a wide audience, which caused indignation among users and analysts.
Meta's press release stated that the Maverick model took second place in ELO rating (1417 points), exceeding GPT-4O from Openai and approaching Gemini 2.5 Pro. These results initially looked impressive and contributed to the positive image of new models. However, users of the Lmaarena platform, where various artificial intelligence models are compared, quickly noticed that the META documentation had an indication that the public version of the model was tested, but a specially tuned experimental version created to optimize the results in the chats.
Meta did not immediately reveal that a customized version of LLAMA-4-MAVERICK-03-26-Expert was used for testing. After this information has become known, the company confirmed that the statistics used for advertising does not display the real characteristics of the public version of models available to users. This has caused criticism from the community that considers such manipulations inadmissible.
The Lmaarena management, in turn, accused META of not complying with the standards of honest testing, which undermined the confidence in their platform as an indicator of real productivity of AI models. In response to the scandal, Lmaarena has already announced his intention to change her rules to avoid such situations in the future. They noted that suppliers who can provide specially optimized versions of their models for testing and then release other public versions, violate the transparency and objectivity of testing.
