Overview
- Meta's Llama 4 models, Scout and Maverick, were released in a surprise weekend launch, featuring a Mixture-of-Experts (MoE) architecture to improve computational efficiency.
- Claims of a 10-million token context window for Llama 4 Scout have been challenged, with real-world tests revealing significantly lower usable limits due to resource constraints.
- Meta has been accused of using an optimized, unreleased version of Llama 4 Maverick for benchmark tests, raising concerns about transparency and the reliability of reported performance metrics.
- Users and researchers have reported inconsistent outputs, repetitive responses, and poor performance on benchmarks, highlighting technical issues with the models' implementation.
- Meta has defended the release, attributing performance variability to stabilization issues rather than intentional manipulation, while pledging ongoing improvements and community collaboration.