How Meta's Llama 3 will be integrated into its AI assistant

With the release of two small open-source models ahead of a major Llama 3 launch later this year, Meta also vows to make the AI readily available across all its platforms.
Written by Don Reisinger, Contributing Writer
The Meta logo outside a building.
Bloomberg/Getty Images

Meta is upping the ante in the artificial intelligence race with the launch of two Llama 3 models and a promise to make Meta AI readily available across all its platforms.

The tech giant on Thursday released two small Llama 3 models ahead of a major Llama 3 launch later this year. The open-source models, which Meta said last week were nearing release, are being integrated into its Meta AI assistant and will be made available to developers.

Also: The best AI chatbots: ChatGPT isn't the only one worth trying

The company told The Verge that Llama 3 was trained on 15 trillion tokens, or bits of information, compared to 2 trillion tokens in last year's Llama 2. The biggest of the two models released on Thursday has 70 billion parameters, or variables it uses to refine results.

That should translate to dramatically improved AI performance compared to Llama 2. Meta is also arguing that the final build of Llama 3 could become the most sophisticated AI option on the market. Even in the small models, Meta has promised better performance in multi-step processes and enhanced performance on complicated queries.

That would be good news for developers who took issue with Llama 2's sub-par performance compared to alternatives from Anthropic and OpenAI. In an interview with Reuters, Meta acknowledged those problems and said that it addressed them by using "high-quality data" as well as AI-generated data to address any problem areas.

Also: Meta's 'pruning' of Llama 2 model shows path to slimmer AI

Still, it's the upcoming major Llama 3 release that could prove most important to developers and Meta itself. While the company has been coy about its final Llama 3, Meta confirmed that it's still training on data and when complete, will have over 400 billion parameters, making it more than five times larger than Llama 2. Unlike the smaller Llama 3 models, the big model will be multimodal, allowing it to generate both text and images.

An expanded Meta AI

Last year, Meta launched Meta AI, a free-to-use AI tool similar to OpenAI's ChatGPT. However, the tool was little more than something to try out.

In addition to its Llama 3 announcement on Thursday, Meta said that it's getting serious about Meta AI. For one, Meta AI will now live on a standalone site, where users can input queries for free. Meta is also integrating Meta AI into the search box across Facebook, Instagram, Messenger, and WhatsApp, and will make its way into the News Feed too, Meta said on Thursday. The company also said that it's expanding Meta AI's availability to several more countries, including Australia, Canada, Jamaica, Nigeria, and Zimbabwe.

Also: Meta unveils second-gen AI training and inference chip

Perhaps most importantly, Meta AI is now powered by Llama 3, making it far more efficient at handling tasks, answering queries, and getting answers from the web. Meta AI's image-generating feature Imagine has also been updated to create images more quickly.

Both Llama 3 and the Meta AI updates are available now. Meta said that it will roll out additional Llama 3 models over time, culminating in the major Llama 3 launch in the coming months.

Editorial standards