Today marks a significant moment in AI development with the launch of Llama 3, the latest large language model (LLM) from Meta Platforms, the parent company of Facebook, Instagram, WhatsApp, and Oculus VR. This model is touted as one of the most powerful open-source AI systems available, debuting shortly after an early leak on Microsoft’s Azure cloud service.
The initial Llama 3 lineup includes two versions: an 8 billion-parameter model and a 70 billion-parameter model, which refer to the complexity of connections between artificial neurons. Meta is also actively training a 400 billion-parameter version, although a release date has not yet been announced. According to Ragavan Srinivasan, Meta’s VP of Product, the performance of these models is exceptional, showcasing benchmark capabilities that rival those of major competitors.
Currently, the Llama 3 8B and 70B models demonstrate performance surpassing or significantly outpacing competing models from Google (Gemini Pro 1.5), Anthropic (Claude 3), and Mistral (7B Instruct) in various tasks. Specifically, the Llama 3 excels in multiple-choice questions and coding tasks but falls short in specific areas like math word problems and graduate-level quizzes. However, the 8B version consistently outperforms other models on grade school math questions.
Meta’s Llama 3 not only makes waves in benchmarks; it is also the driving force behind a new standalone Meta AI chatbot available on the web at www.meta.ai. This chatbot directly competes with OpenAI’s ChatGPT, Anthropic’s Claude 3, and HuggingFace’s HuggingChat. As Mark Zuckerberg stated, “Meta AI is the most intelligent AI assistant that you can freely use.”
The Meta AI chatbot offers a user-friendly dialog interface and can be accessed without a Facebook login, though functionality may be limited initially. It includes an integrated image generation model, Meta Imagine, which allows for near real-time creation of images based on user prompts. Users will need to log in with their Facebook accounts to utilize this feature.
Unlike some other platforms, the Meta Imagine generation does not currently allow for aspect ratio adjustments for generated images but offers a unique watermark indicating AI generation. Users can see the image creation process live and replay it as an animation or GIF.
Meta AI also sources live information from both Microsoft’s Bing and Google’s search results, enhancing its capability to deliver up-to-date responses. However, the current version of Llama 3 lacks multi-modal functionality, meaning users cannot upload images or documents for analysis, but Meta plans to address this in future updates.
A key distinction of Llama 3's open-source status is that it operates under a Meta-specific license, similar to its predecessor, Llama 2. While third-party enterprises can use Llama for commercial purposes, those with over 700 million monthly active users must request a license from Meta. This has drawn criticism from the open-source community as it may limit competition.
Regarding improvements, Llama 3 offers reduced false refusal rates, better alignment, and greater diversity in responses compared to Llama 2. This enhancement stems from a comprehensive training process utilizing a three-pronged parallelization strategy, which increased training efficiency by threefold. Llama 3 has been trained on over 15 trillion tokens, a sevenfold increase from Llama 2.
The 70B version also features an 8,000-token context window, nearly double that of its predecessor, allowing for longer prompts and more comprehensive information input. Llama 3 is currently available for download on Meta's AI website and is being hosted on platforms like AWS, Google Cloud, and Microsoft Azure.
For more information on Llama 3, visit Meta's official AI website.