LlamaMeta's open family of large language models — released weights that sparked an open-source AI wave. 4 brought mixture-of-experts and native multimodality to Meta's open model line. Scout fit on a single H100; Maverick beat GPTGenerative Pre-trained Transformer — OpenAI's family of large language models from GPT-1 through GPT-5.-4o and GeminiGoogle's family of multimodal AI models — successors to PaLM and competitors to GPT and Claude. 2.0 FlashAdobe Flash — a browser plugin for animations and video — deprecated as the web moved to HTML5. on several benchmarks. Meta also previewed the training-only Behemoth teacher — signaling open weights would keep pace with closed frontier models.

LlamaMeta's open family of large language models — released weights that sparked an open-source AI wave. 4 made open MoE multimodalAI models that process multiple input types — text, images, audio, and video in one system. models competitive with closed APIApplication programming interface — a defined way for programs to talk to each other or to a service. flagships.

Ten-million-token context and MoE efficiency reset expectations for self-hosted AI.