Meta has released Llama 3, its most advanced large language model (LLM) to date. The social media giant confirmed that the AI model, which it is making open source, would be accessible via AWS, Google Cloud, IBM WatsonX and other platforms. Meta said that Llama 3 showcases enhanced reasoning capabilities compared to its predecessors, which it largely ascribes to improvements in its transformer architecture, training regime and instruction tuning. While the firm characterised the model as “early,” the model is already connected to the firm’s virtual assistant, Meta AI. The assistant is not currently available in the UK.
“We are embracing the open source ethos of releasing early and often to enable the community to get access to these models while they are still in development,” said Meta. “The text-based models we are releasing today are the first in the Llama 3 collection of models. Our goal in the future is to make Llama 3 multilingual, have longer context, and continue to improve performance across core LLM capabilities such as reasoning and coding.”
Llama 3 showcases improvements in training
Llama 3 exists in both 8bn and 70bn parameter formats. Improvements in the model’s training, claimed Meta, led to superior performance in human evaluation tests against other LLMs like Claude Sonnet, Mistral Medium and OpenAI’s GPT-3.5 (data on how it compared to the latter’s successor, GPT-4, was not included in Meta’s announcement.) The firm attributed Llama 3’s success in these trials to the creation of new performance benchmarks to “optimize for performance for real-world scenarios,” including an evaluation set containing 1,800 prompts covering use cases including brainstorming, classification, creative writing and answering open questions.
Llama 3 is also equipped with new coding abilities and “improved reasoning,” said Meta. Though only capable of producing text outputs, the firm explained that new features – including features like multimodality and “the ability to converse in multiple languages” – will be added in due course. Open-sourcing the model will also allow developers to chime in on fine-tuning the model and suggest improvements, said Meta. “We want to kickstart the next wave of innovation in AI across the stack,” it explained, “from applications to developer tools to evals to inference optimizations and more.”
Meta open source commitment questioned in past
Meta also open-sourced Llama 3’s predecessor Llama 2 after its release in 2023, though some experts criticised the firm for failing to live up to this promise by imposing limits on developers from its competitors experimenting with the LLM. Some users also said the model had too many safety filters and responded to requests to write songs or change file names with long screeds about political correctness. Llama 2’s reputation was further tarnished by Meta’s president of global affairs, Nick Clegg, who told BBC Radio 4’s ‘Today’ programme that most LLMs were “quite stupid.”
Victor Botev, chief technology officer at Iris.ai, welcomed Meta’s release of Llama 3 as an opportunity to showcase the value of open-sourcing models in AI model development. “However, it’s important to maintain perspective – a model simply being open-source does not automatically equate to ethical AI,” said Botev. “While open initiatives like Llama 3 promote scrutiny and collaboration, their true impact hinges on a holistic approach to AI governance compliance and embedding ethics into AI systems’ lifecycles. Meta’s continuing efforts with the Llama model is a step in the right direction, but ethical AI demands sustained commitment from all stakeholders.”
Read more: Alibaba brings Meta’s Llama 2 AI model to China
