Meta has unveiled the latest version of its large language model (LLM), known as Llama 3. The company announced the launch on April 18, touting it as a significant improvement over its predecessor, Llama 2.
According to Meta, the initial release of Llama 3 includes models with 8B and 70B parameters, with future versions expected to feature a whopping 400B parameters. The new model was trained on a vast dataset of over 15 trillion tokens, which is seven times larger than the training data used for Llama 2.
One key feature of Llama 3 is its filtering techniques, including NSFW filters, to ensure data quality. The company claims that Llama 3 outperforms competing models like Anthropic’s Claude Sonnet, Mistral Medium, and OpenAI’s Chat GPT-3.5 in a variety of use cases.
While the initial releases of Llama 3 are text-based, Meta plans to introduce multilingual and multimodal versions in the future. These models will offer longer context and improved performance in reasoning and coding, which Meta considers essential capabilities for large language models.
In addition to targeting developers, Meta is making AI services more accessible to end users in the US and over a dozen other countries. The company has launched a new website called Meta AI, where users can access AI-powered writing assistance, trivia games, job interview simulations, and homework help.
Meta has also integrated Meta AI with its various products, including Facebook, Instagram, WhatsApp, and Messenger. The company is even offering AI services through its Ray-Ban Meta smart glasses in the US and plans to expand this to its Meta Quest VR headset.
CEO Mark Zuckerberg expressed confidence in Meta AI, claiming it to be the most intelligent AI assistant available for free use. This announcement comes on the heels of upgrades to competing services like ChatGPT and Microsoft Copilot, signaling a growing competition in consumer-oriented AI services.
