By Roman Aglaganov
In February 2023, Meta released LLaMA, a family of language models, to approved researchers. About a week later, on 3 March, the model weights were posted as a torrent, the link was shared on 4chan, and within days anyone could download them. It was not the launch strategy anyone had planned, but it changed the industry.
From leak to strategy
Instead of fighting the tide, Meta leaned in. In July 2023 it released Llama 2 openly, with Microsoft as a partner and a licence allowing commercial use for almost everyone (companies with more than 700 million monthly users needed a separate licence, a clause aimed squarely at Meta’s biggest rivals). Llama 3 followed in April 2024, and in July 2024 Llama 3.1 included a 405-billion-parameter model that competed with closed leaders.
Open weights, not quite open source
Llama models are open-weight: you can download and run the trained model, modify it and build on it. Purists note that is not the same as open source, because the training data and full recipe are not published and the licence has conditions. Either way, thousands of developers fine-tuned Llama for specific languages, industries and tasks.
Why Meta gives it away
Meta’s argument is that open models become industry standards, attract talent and stop rivals from owning the platform everyone builds on. Not everything went to plan: Llama 4, released in April 2025, got a lukewarm reception, and rivals such as DeepSeek, Alibaba’s Qwen and Mistral now compete hard in open-weight AI. But the movement Llama kicked off is very much alive, as is the llama pun industry.
