Meta releases Llama 2 LLM, free for commercial use

'More collaboration, more scrutiny, more innovation is good for the company, good for the business environment, good for startups, good for policymakers,' says director of public policy Chris Yiu

Meta has released a version 2 of it's Llama large language model (LLM) under an open licence, meaning it can be used for free (with some caveats), including for commercial purposes.

Use of the original Llama was restricted to researchers.

Llama 2 is a pre-trained LLM which has also been refined using reinforcement learning with human feedback (RLHF). The training data contained 40% more tokens than the original model, according to Meta, and RLFH, which is also used by OpenAI and Google in an effort to stop chat products ChatGPT and Bard overstepping boundaries, is a safety measure.

Indeed, Meta researchers describe Llama 2 as "optimised for dialogue" leaving little doubt that the company is entering the AI chat race.

The licencing arrangement sets it apart from GPT-4 and Google PaLM 2 (the models on which ChatGPT and Bard are based), as those models are not free for commercial use.

Llama 2, released in partnership with Microsoft Azure, comes in three sizes 7B, 13B and 70B, the B standing for billions of parameters in the training data.

The models can be downloaded for free, but there are restrictions on their use. They cannot be used to improve other AI models, for example, or used with applications with more than 700 million monthly users.

As such, unlike LLMs such as Falcon and MPT which use the Apache 2 licence, the Meta community licence is not approved by the Open Source Institute and so Llama 2 cannot be officially considered open source.

That quibble aside, Meta's emphasis on transparency and openness is to be welcomed, said Amanda Brock, CEO of non-profit advocacy group OpenUK.

"The democratisation of AI and opening innovation around it to enable collaboration across our open communities is an essential step in the future of this most impactful of technologies," she said.

"Whilst we recognise that this licence is not an official ‘open source software' one, it provides certainty for those that want to take advantage of Llama 2. This is a fantastic step in the right direction for open AI communities."

Speaking at an OpenUK press event to launch a research report into openness and AI last week, Chris Yiu, director of public policy at Meta, said opening up AL models is important for democratising control. "I think we would like to see much more collaboration between businesses, startups, government, academia, civil society, there should be an effort where actually we will have a stake in this, we all have something to contribute."

Transparency will help to overcome some public concerns about AI, he added: "More collaboration, more scrutiny, more innovation is good for the company, good for the business environment, good for startups, good for policymakers."

Openness also helps Meta find flaws develop new products more quickly, he said, in case anyone was about to accuse the company of altruism.

However, Yui said he didn't believe all AI models should be open: "There will be reasons that some things to remain proprietary, and I think that's fine. There should be a healthy mix."