Meta Open-Sources Llama 4 Architecture, Keeping Weights Closed

In a controversial pivot from its entirely open-source strategy with Llama 2 and 3, Meta has announced the architecture for its upcoming Llama 4 model will be open-sourced, but the actual trained weights will remain heavily restricted. Mark Zuckerberg cited the immense capabilities of the new model, specifically regarding autonomous hacking and biosecurity risks, as the primary reason for the restriction.

This hybrid approach has deeply fractured the open-source AI community. While researchers appreciate access to the advanced architectural papers detailing novel attention mechanisms and MoE (Mixture of Experts) routing, the lack of weights severely limits the ability of the broader community to fine-tune and deploy the model independently.

The Shifting Definition of 'Open AI'

Meta's decision highlights the growing anxiety surrounding highly capable AI models. By keeping the weights closed, Meta ensures they remain the sole gatekeeper of Llama 4's deployment, allowing them to implement strict safety guardrails and monetize enterprise access via API, mirroring OpenAI's strategy.

Impact on Independent Research

Independent AI researchers and startups heavily reliant on Meta's previously open ecosystem will be forced to look elsewhere, potentially accelerating the adoption of alternative open-weight models from platforms like Mistral or newly emerging decentralized AI collectives.