Llama 4 Scout

Llama 4 Scout is Meta's first MoE release, and the public reception was mixed enough to be a story in itself. The headline features were a 10M-token context window (the longest in any frontier model at release) and native multimodal input, both delivered alongside an architecture pivot that followed DeepSeek V3 in establishing MoE as mainstream for open-weights frontier work. Independent evaluators noted that the announced benchmark scores were not consistently reproducible at deployment, that the 10M-token context was achievable in theory but degraded at lengths well below the stated maximum, and that the LMArena ranking the launch material featured was for a chat-tuned variant not available as weights. The release itself remains historically significant as the moment Llama abandoned dense scaling, but the immediate developer narrative was that DeepSeek V3 and Qwen 3 had executed the open-weights MoE story more cleanly five months earlier. Llama 4 Scout's lasting value depends on whether the Maverick and Behemoth siblings shipped on the same architecture deliver on the long-context and multimodal promises in production deployment.