What Open-Source Means in This Context
Llama 3 is not open-source in the traditional software sense β the training data is proprietary, and the license restricts certain commercial uses. But the weights are publicly available, which means any developer with sufficient compute can run, fine-tune, and deploy the model without paying Meta.
That distinction β weights public, training data private β is the new frontier of AI openness, and it represents a genuine shift in the power dynamics of the industry.
Why Meta Did This
The cynical reading is that Meta does not have a compelling AI product to sell, so open-sourcing the model commoditises a layer where its competitors (OpenAI, Anthropic, Google) charge significant API fees. Commoditising your competitor's moat is a classic strategic move.
The more generous reading is that Meta genuinely believes open AI development produces better outcomes β more researchers improving the model, more diverse applications, more public scrutiny of capabilities and risks.
The Downstream Effects
Within weeks of Llama 3's release, hundreds of fine-tuned variants had appeared on Hugging Face. Developers were running capable AI models on consumer hardware. The inference cost for frontier-quality responses dropped by approximately 90% compared to GPT-4-class API pricing.
For startups building on top of AI, this changes the build-vs-buy calculation fundamentally.