Benchmark’s brilliant results of Lama 4 meta artificial intelligence models were questioned

Meta unveiled the Llama 4 artificial intelligence models earlier this week, including Scout and Maverick. The results of the initial benchmarks showed that these models were stronger than competitors, but the meta may have manipulated the results of these benchmarks, though the company itself rejects the claim.

According to Verj, the Maverick model in the LMARNA benchmark was able to score 2 and pass competitors such as the 4O model of Openai and reach a very short distance behind the Jina 4.0 Pro. The results initially indicated that Meta was able to defeat powerful competitors such as Google and Openai, but artificial intelligence experts examined meta claims and achieved interesting results.

Doubt about the results of the LLAMA 4 benchmarks

Experts claim Meta to teach LLAMA 4 models to perform better in benchmarks and hide their real restrictions. Although there were a lot of doubts about the performance of these models in benchmarks and the possibility of manipulating the results, Ahmad al -Dahlah, the deputy director of artificial intelligence in Meta, denied all the rumors in a post on the X -Social Network, saying it was not true.

Meta acknowledges a version of the Maverick AI tested in the LMARNA benchmark, not the same as the public. According to Meta’s own documents, the company has put a trial version of the Maverick model in the Lmarena benchmark. This model is optimized for conversation and uses more emoji. It also gives longer answers to users’ questions.

In addition, the timing of the release of new models of meta artificial intelligence added to the experts. Meta unveiled the Scout and Maverick models on Saturday, the first day of the weekend, which is very unusual to unveil a new product with such extraordinary advances.