Llama 4
Meta moved the whole open line to a mixture of experts and gave the smaller model a context window of ten million tokens.
Why it matters
The open line changed architecture entirely, and at the same time failed for the first time to keep the promises it made on the measures.
Scout holds 109 billion parameters with 17 active, Maverick 400 billion with 17 active. The largest of the family, Behemoth, stayed in preview. The release came with a dispute: the version submitted to the human-preference ranking differed from the published weights, which Meta's own post called an experimental chat version and the ranking's operator called a customized model optimised for human preference. The episode showed what public measures are worth when a vendor can submit a different model to them, and the ranking changed its rules afterwards.