Meta announced the release of Llama 4 405B, its largest open model to date. With 405 billion parameters, it becomes the most capable open-source model ever released, reaching 87.5% on MMLU and 89.0% on HumanEval — scores that place it in the same neighborhood as GPT-4.5 and Claude Sonnet 4.
The Llama Community license has been slightly broadened compared to version 3, now allowing commercial use up to 1 billion monthly active users without specific negotiation (previously 700 million). Restrictions still concern use for training other commercial models.
The model is available for direct download, via Hugging Face and AWS Bedrock. Meta also releases 4-bit and 8-bit quantized versions that allow running the model on a single 8×H100 machine, making its use feasible for medium-sized tech teams.