New Llama generations improved multimodality and agents while keeping downloadable weights as industrial strategy.
Meta continued releasing Llama 3.x and progress toward Llama 4 with stronger multimodal and tool-use abilities, feeding startups and clouds.
The strategy: commoditize models to sell hardware, ads, and surrounding services. Developers gained strong baselines without API lock-in.
The open-weight ecosystem became infrastructure as much as any single vendor’s product.
Meta released LLaMA and, after a leak, catalyzed an open fine-tune ecosystem that challenged closed APIs.
Meta, Mistral, DeepSeek, and others brought public weights near closed-API performance, redefining “frontier.”
Meta Superintelligence Labs shipped Muse Spark 1.1 with Thinking mode on meta.ai and a public preview of the Meta Model API, nearly a year after Llama 4.