Meta is about to launch its biggest Llama model yet — here's why it's...
Meta recently unveiled Llama 3, the latest addition to its AI-powered large language models lineup. With a dataset that surpasses Llama 2 by at least seven times, Llama 3 initially debuted in 8B and 70B parameter sizes. Impressively, Llama 3 outperformed not only Llama 2 but also Google's Gemma and Anthrophic’s Claude Sonnet upon its launch. Following Sonnet's upgrade, it emerged as one of the most potent AI models available.
Now, leaks hint at the imminent launch of the eagerly anticipated Llama 3 models, which have been trained on over 400 billion parameters. This significant development is just one of several new models that Meta is diligently working on, leveraging its vast resources, including thousands of Nvidia H100 GPUs.
Technical Advancements and Efficiency
In initial testing, the instruction-tuned Llama 3 400B achieved an impressive score of 86.1 on the MMLU benchmark, placing it on par with GPT-4's performance while utilizing less than half the parameters. This achievement underscores Meta's strides in model architecture and training, positioning them to challenge the dominance of OpenAI.
Large language models with more parameters typically deliver superior performance on benchmarks and real-world tasks. The fact that Llama 3 400B can almost match GPT-4's MMLU score with fewer parameters indicates enhanced efficiency in terms of computational resources, energy consumption, and cost compared to OpenAI's ChatGPT 4.
Open License and Accessibility
One compelling aspect of Llama 3 is its release under an open license for both research and commercial purposes. While it remains uncertain if the 400B model will follow suit, an open release would make state-of-the-art language capabilities accessible to researchers and developers through various cloud platforms and ecosystems, fostering innovation and diverse technology applications.
The availability of the 400B model, equipped to rival ChatGPT 4, empowers researchers to expedite the development of advanced language AI applications independently, without reliance on costly proprietary APIs.
Meta AI has been teasing the release of the 400B model since the unveiling of Llama 3, promising new capabilities such as multimodality, multilingual conversation, extended context windows, and enhanced overall performance. While speculation abounds regarding the release date for the 400B models, Meta has confirmed the completion of development without a set launch date.
Anticipated Release and Beta Testing
WhatsApp Beta users on Android 2.24.14.7 have noted a new feature allowing them to experiment with the Llama 3-405B model from Meta AI. Although this option is currently limited to beta users and restricted in usage volume, it signals a potential full release, expected around late July or August 2024.
As excitement mounts within the tech community, the prospect of Meta's groundbreaking Llama 3 models underscores a new era of possibilities in AI development and application.
