In mid-July 2023, Meta released Llama 2, a family of large language models, with a licence that permitted commercial use for organisations under approximately seven hundred million monthly active users. The release was a significant change from the original Llama, which had been distributed only for research and had then leaked publicly, leading to widespread but legally ambiguous use. Llama 2 was the same kind of model with explicit permission to build on top of it.
The technical capabilities of the released models, particularly the seventy-billion-parameter variant, were broadly competitive with the GPT-3.5 class of models that had dominated the previous twelve months. The benchmarks were not as strong as GPT-4. They were good enough to be a serious option for many production workloads, particularly where the cost or control advantages of running an open model outweighed the capability gap with the proprietary frontier.
The licence terms mattered enormously. With commercial use permitted under reasonable conditions, organisations that had been hesitant to invest engineering effort in fine-tuning and deploying open source language models had clear green lights. The fine-tuning ecosystem expanded rapidly. Specialised variants of Llama 2 for particular domains and tasks started appearing within weeks. The serving infrastructure improved as more demand created more attention.
The strategic logic for Meta was layered. Open releases of strong models damaged the moat of competitors who relied on proprietary access to their models as a differentiator. They also positioned Meta as the centre of an open AI ecosystem, with influence over standards and development practices that the company would not have if the field consolidated around closed models. The cost of training the models was significant, but for a company with Meta’s revenue base, the cost was tractable in exchange for the strategic positioning.
The open source AI conversation changed substantively after Llama 2. Discussions of whether open source models could ever be competitive with proprietary ones largely shifted to discussions of how the gap would evolve over the following year and which categories of use would prefer which type of model. The argument had been settled in favour of open source being a serious option, even if the question of whether it would catch up to the frontier remained open.