Meta’s Llama 4 Model Different in the Test, Really Different!

Meta, New artificial intelligence models He introduced it, but then things got a little confused. The company’s models rmore successful It was claimed that he played a little in his favor ”. In short, Meta’s artificial intelligence seemed smarter, but there was a small game behind it.

From the company’s Llama 4 series Maverick ModelsAlthough it received high scores in the tests, it turned out that this result did not reflect the public version.

The model presented to the public is not the same in the test.

MetaWith its announcement at the weekend Llama 4 introduced two new artificial intelligence models under the roof. These models called Scout and Maverick, especially Maverick, GPT-4O and Gemini 2.0 flash, such as competitors said. Maverick rising to the second place on the popular comparison platform called LMarena, drew attention with 1417 ELO points.

However, the details that emerged soon showed that the tested Maverick was different from the public version. In the documents of the Meta, the model used in LMarena is a “an experimental version that has been specially optimized for speech skills”. Patform managers also reacted and announced that the rules will be updated.

Meta, while defending the use of this special version, the open -source version of developers can customize according to their needs, he said. However, the different model presented to the public with the performance shown on the test platform led to the questioning of the yield that can be obtained in the real world. The company’s model is also known to postpone several times before.

This development once again brought the importance of artificial intelligence comparison tests in terms of transparency and reliability. At the same time, the difference between the test conditions of the models and the user -presented state may also affect developers’ decision -making processes.