AI news
Meta Releases Llama 4: The First Open--Weights 'World Model'
Meta has officially launched Llama 4, a 1.5--trillion parameter model family designed with a native "World Model" architecture. This shift allows the model to process and predict physical interactions and temporal sequences in video with high fidelity, moving beyond standard text--based reasoning.
| Metric | Detail |
|---|---|
| Model Size | 1.5T (Flagship), 400B, 70B |
| Context Window | 2.0M Tokens |
| Training Compute | H100/B200 Cluster (Estimated 10^26 FLOPs) |
Strategic Implications:
- Open--Source Leadership: By releasing the weights, Meta effectively commoditizes the "Frontier" model layer, pressuring the margins of closed--source competitors.
- Robotics Integration: The model’s ability to understand physical causality makes it a turnkey solution for humanoid robot manufacturers seeking high--level reasoning kernels.
- Expert Commentary: Analysts suggest Llama 4's efficiency in "System 2" thinking (deliberative reasoning) reduces the need for external RAG pipelines in complex legal and scientific workflows.