Llama 2: open(ish) weights change the game (2023)
Meta released a capable language model whose weights anyone could download and build on, pushing the debate over 'open' AI to the center of the industry.
In July 2023, Meta released Llama 2, a family of large language models in several sizes, together with a chat-tuned version. What set it apart was not a benchmark score but the licence: the model weights were made available to download, free for research and, within limits, for commercial use.
What “open weights” means
Models like GPT-4 are reached only through an API. You send text in, you get text back, and the actual parameters stay on the provider’s servers. Llama 2 was different: you could obtain the weights and run the model on your own hardware, inspect it, and adapt it.
That unlocked a lot:
- Fine-tuning. Teams could fine-tune Llama 2 on their own data to specialise it, without sending that data to a third party.
- Privacy and control. Running locally meant sensitive prompts never left an organisation’s own systems.
- A shared base. A common, capable starting point let a whole ecosystem of variants and tools grow on top of it.
Llama 2 was not fully “open source” in the strict sense. Its community licence carried conditions, including limits on the largest platforms, so many people prefer the more precise term open weights.
The ecosystem effect
Access mattered as much as the model itself. A wave of fine-tuned derivatives, tools, and research followed, much of it from people who could never have trained a model of that scale from scratch. The chat version was refined with human feedback, the same RLHF idea behind ChatGPT, giving the community a usable assistant to build on rather than just a raw base model.
Why it matters
Llama 2 crystallised one of the defining tensions of the era: closed frontier models versus open weights. One camp argues that the most powerful models should be tightly controlled for safety and commercial reasons; the other argues that open weights spread benefits, enable scrutiny, and prevent a handful of companies from owning the technology.
That debate is still unresolved, but Llama 2 is the moment it became unavoidable, proving that openly available models could be good enough to matter.