Meta today released Llama 3, the next generation of its large language model, available in 8 billion and 70 billion parameter sizes. Meta says the release is industry leading. A 400 billion parameter version is still training and expected in the coming months.

The models are released under a license, and Hacker News commenters quickly point out that it is not truly open source. One commenter called it ‘open weight, not open source,’ while another emphasized the need to use terms like ‘weights-available’ to avoid open-washing. Others defended Meta’s approach as a good balance that enables fine-tuning and deployment without demanding full reproducibility of training.

The community is buzzing with speculation about the impact. Some see it as a threat to OpenAI and Google, while others expect it to benefit the broader ecosystem by lowering barriers. One commenter links a Colab notebook for LoRA fine-tuning Llama 3 8B. The lack of comparisons to GPT-4 and Claude Opus in Meta’s announcement has not gone unnoticed, but the promise of a future 400B model keeps the conversation lively.