Meta today released Llama 3.1 405B, a 405-billion-parameter model it says is the first open-weights AI model to match the capabilities of leading proprietary systems like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet. Trained on over 15 trillion tokens using 16,000 Nvidia H100 GPUs, the model is text-only but achieves competitive results across coding, math, multilingual translation, and general knowledge benchmarks.

Alongside the flagship, Meta refreshed its 8B and 70B models with 128,000-token context windows — a 16x increase over the prior Llama 3 generation — and added support for eight languages and tool use. The company also updated its license to explicitly allow developers to use outputs from Llama models to train other models, which Meta says helps with synthetic data generation and model distillation, addressing a major criticism from the open-source community. However, the license still restricts developers with over 700 million monthly active users, who must seek special permission.

In a simultaneous open letter, CEO Mark Zuckerberg argues that open source is good for developers, Meta and the world, and frames the release as part of Meta’s open-access push. The release comes with a reference system, safety tools (Llama Guard 3, Prompt Guard), and a call for comment on a proposed Llama Stack API. More than 25 launch partners — including AWS, Nvidia, Databricks, Groq, and Google Cloud — announced support on day one.

The move intensifies the open-vs-closed AI debate. Critics note that the model still exhibits hallucinations, requires a server node for inference, and trails Claude 3.5 Sonnet on some reasoning tasks. But for many developers, the offer of a frontier-level model they can download, fine-tune, and deploy without sharing data with Meta is the strongest case yet for open-source AI’s viability.