Meta has rolled out its latest innovation, Meta Llama 3, marking a significant advancement in open generative AI with large language models (LLMs). The tech giant has unveiled two variants of the model, boasting parameter counts of eight billion and a staggering 70 billion, respectively, promising versatile applications across various domains. Meta asserts that these models represent the pinnacle of open-source AI within their category, emphasizing their superior capabilities.
Looking ahead, Meta envisions enhancing these text-based models with multilingual and multimodal support, along with continuous improvements in coding proficiency and reasoning abilities. Notably, the training dataset for Llama 3 dwarfs its predecessor, Llama 2, by a factor of seven, underscoring Meta’s commitment to refining model performance. To ensure the quality of training data, Meta has deployed a suite of data-filtering pipelines incorporating heuristic filters, NSFW filters, semantic deduplication techniques, and text classifiers.
Moreover, Meta has integrated robust safety features such as Llama Guard 2, Code Shield, and CyberSec Eval 2 into the new model, underscoring its commitment to user protection. The social media behemoth has outlined plans to make Llama 3 available across major platforms, including cloud service providers and model API providers, among others.
Meta’s integration of the latest Llama models into Meta AI extends their accessibility to users across its ecosystem of apps, including Instagram, Facebook, WhatsApp, Messenger, and web platforms. Additionally, users can experience the enhanced capabilities of multimodal Meta AI through Meta’s innovative Ray-Ban Meta smart glasses.
The widespread availability of Llama 3 models is set to extend beyond Meta’s ecosystem, with plans for integration into prominent platforms such as AWS, Google Cloud, and Microsoft Azure, among others, bolstered by support from leading hardware providers like AMD, Intel, and NVIDIA. Meta has also hinted at the development of even larger models, exceeding 400 billion parameters, with forthcoming releases slated to introduce advanced features such as multimodality, multilingual conversational abilities, expanded context windows, and overall enhanced performance.