Llama 4 Released: Open Models Shift to Native Multimodal MoE
Meta released Llama 4 Scout and Maverick, adopting native multimodal and mixture-of-experts architecture.
Open model roadmap follows multimodal, ultra-long context, and sparse activation, continuing to narrow the fundamental capability gap with closed-source models.
MoE decouples total parameter count from inference cost; Scout extends context to tens of millions of tokens.
Open ecosystem enters complex system competition, but training transparency, evaluation credibility, and deployment barriers remain constraints.
Privatization and open-weight options become richer; enterprises need to assess total cost of ownership rather than just benchmarks.
Observe community reproduction, actual context utilization, enterprise deployment, and subsequent flagship models.