On April 18, Meta made the announcement that Llama 3, its most recent large language model (LLM), had launched. It was hailed as a “major leap over Llama 2.”
According to the company, it has already released the first two models of the current version, which have 8B and 70B parameters. 400B parameters will be featured in future models.
A “large, high-quality training dataset” with over 15 trillion tokens—7 times larger and 4 times more code than Llama 2—was used to train Llama 3, as highlighted by Meta. To maintain the quality of the data, Llama 3 also includes filtering methods, such as NSFW filters.
Over half of the 12 use cases show that LLama 3 performs better than Llama 2 and rival models like Claude Sonnet from Anthropic, Mistral Medium, and Chat GPT-3.5 from OpenAI.
Text-based models comprised the initial releases of Llama 3. But multilingual and multimodal releases are on the way. “Core LLM capabilities” as defined by Meta will be exhibited by them, along with a longer context and improved reasoning and coding performance.
All significant cloud providers, model API providers, and other services will host Llama 3, according to the company’s plans. The product will be released “everywhere,” as planned.
Greater user Accessibility
Developers are the target audience for Llama 3, but Meta has also introduced new channels for end users in the US and over 12 other countries to access AI services.
A recent inclusion is a specialized website called Meta AI, where users can get homework help, trivia games, simulated job interviews, and writing help powered by AI.
Facebook, Instagram, WhatsApp, Messenger, and other products from Meta are all integrated with Meta AI. Additionally, the service is available in the US through Ray-Ban Meta smart glasses, and the company has plans to expand it to include its Meta Quest VR headset.
The announcement of Meta’s expanded AI product line follows the release of updates to rival services. The competition between consumer-focused AI services progressed when ChatGPT upgraded to GPT-4 Turbo on April 11 and Microsoft Copilot upgraded to GPT-4 Turbo beginning in March.