Meta has finally announced the release of Llama 3, the latest iteration in the company's series of Llama large language models. This new model comes with enhanced capabilities and is available in two configurations: 8 billion and 70 billion parameters, each designed to cater to a broad spectrum of AI applications. It introduces major improvements, including more robust reasoning skills, and promises a significant enhancement in its coding and instruction-following abilities, promising to be the most versatile and capable open-source model available to date.

Meta has emphasized its intent to continue promoting innovation in AI by making these powerful tools accessible for broad use. This approach is expected to give power to developers, researchers, and businesses to push the boundaries of what AI can achieve, as it can run locally on all kinds of machines. The 8B parameter model is particularly lightweight, but early reports suggest that it almost beats out Wizard 8x22B, a 140B Mixture of Experts model.

Goals and Innovations of Meta's Llama 3

It outperforms a ton of different models

With Llama 3, Meta aims to match or even surpass the capabilities of the best proprietary models currently on the market. The model addresses previous feedback from developers by enhancing overall usefulness while maintaining a focus on responsible deployment. Llama 3 is designed to be multilingual and multimodal in future iterations, promising longer context windows and continuous performance improvements. As an example, it outperforms Claude 3 Sonnet according to Meta, and Sonnet already far surpasses GPT-3.5. You can read about how Meta calculated its figures in the company's evaluation document.

The technical enhancements in Llama 3 include a more efficient tokenizer and a novel approach to model training that uses grouped query attention, optimizing the model for better inference performance. The improvements in Llama 3 aren't just of a technical nature either but aim to refine the model's interaction with users, making it more intuitive and effective in day-to-day usage. This was achieved through a training process for Llama 3 that was more extensive and refined than its predecessors. It involves a massive dataset of over 15 trillion tokens, with a significant focus on high-quality, diverse sources. This extensive training, Meta says, is complemented by advanced data-filtering techniques to ensure the quality and relevance of the training material.

Meta has also adopted sophisticated training mechanisms such as data, model, and pipeline parallelization to efficiently manage the vast computational demands of Llama 3. These techniques not only improve the training speed but also enhance the model's ability to learn from an expansive dataset without compromising performance.

With those improvements, Meta also isn't shying away from the ethical questions of AI. The company claims to have integrated extensive testing and red-teaming to identify and mitigate potential misuse of Llama 3. This includes advanced safety features and the new Llama Guard 2 framework, which provides developers with tools to ensure safe and responsible usage of the AI. It remains to be seen what exactly this entails or how it works, but it should hopefully prevent the AI from being misused.

How to use Llama 3 today

It's available for usage right now on Meta AI, and you can also download the model

Llama 3 is beginning its rollout, and Meta is already preparing for future enhancements. These include plans for models with over 400 billion parameters, improvements in multimodal capabilities, and support for multiple languages. The development of Llama 3 is an ongoing process, with Meta committed to regular updates and improvements. You can already try it out today by going to the Meta AI website, and it doesn't even need an account to use it.

If you want to run Llama 3 locally, you can check out Meta's Llama 3 website to get up to speed with its capabilities and how to run it, complete with model downloads for you to try out. You'll likely need some powerful hardware to run the bigger model on your PC, but the 8B parameter model should be easy enough for anyone to run.