Meta releases latest AI model with plans to highlight more on Instagram

A panel at Meta AI Day in London earlier this month. (AP)
A panel at Meta AI Day in London earlier this month. (AP)


The social-media company is unveiling Llama 3, a follow-up to its large language model that will be available to the public

Meta Platforms on Thursday released Llama 3, the latest edition of the company’s publicly available large language model, as the AI battle in Silicon Valley continues to escalate.

The company’s new AI capabilities will be used to upgrade its Meta AI tool, which it will now feature more prominently in its apps. They will also be integrated into the search functions of its major platforms, which include Facebook, Instagram, WhatsApp and Messenger. The upgraded tool will also be able to summarize links. Meta AI will also now be available via the web at

Meta is releasing two versions of the model to start, including an 8 billion parameter model and another with 70 billion parameters. Parameters are the term for the variables in an AI system that is used to measure a model’s size and capabilities. 

The company said the smallest version of Llama 3 is as powerful as the largest version of its predecessor. The company will also be using its Llama 3 model to upgrade its Meta AI assistant chatbot.

“Based on the benchmarks we’re looking at today, this is going to be the most intelligent and best performative model of its kind that’s freely available," Chief Product Officer Chris Cox said in an interview.

OpenAI kicked off Silicon Valley’s AI race in late 2022 when it launched its innovative ChatGPT chatbot. Since then, more companies, both startups and established tech titans, have entered the fray.

In July, Meta publicly launched Llama 2, and it released the first version of its Meta AI chatbot in September within Instagram, WhatsApp and Messenger. Unlike others in the space, Meta launched its model as “open source" software, meaning that others can use it and modify it for free.

Meta said it would release more Llama 3 models over the coming months with additional capabilities, including the ability to converse in several languages. The company is working on an even larger model of 400 billion parameters that it will release after the company has completed its training, Cox said. 

OpenAI hasn’t disclosed the size of GPT-4, but it has been estimated to be 1.5 trillion parameters.

Another key difference between Llama 2 and 3 is that Meta has improved its fine-tuning so that its new model will refuse to answer less-innocuous prompts such as, “How do I kill a computer program?" while continuing to decline to answer prompts including, “How do I kill my neighbor?"

With the launch of Llama 3, Meta AI will be able to do more, and it will roll out to 13 more English-speaking markets, including Australia and Canada. It won’t be available in the European Union at launch.

“It’s going to be a lot more powerful and a lot more intelligent," Cox said. He added that the link summary feature could be helpful for the company’s users when they come across long-form content.

Meta AI’s image-generation feature, called Emu, will also be upgraded to more quickly produce results that will be of higher quality and show users preview results as they type out their query.

When it comes to generative AI, Meta is focused on democratizing these AI breakthroughs for free for its users, Cox said.

“What we’ve tried to do as a company is make the tools work for everyone everywhere," Cox said. He added that the company would consider a monthly subscription model like OpenAI’s ChatGPT Plus and Google One AI Premium when larger versions of Llama become available.

Notably, Meta used synthetic data, meaning text generated by other AI models, to train Llama 3. As tech companies run out of data with which to train their AI models, they are having to come up with solutions for how to acquire or create more data.


Catch all the Corporate news and Updates on Live Mint. Download The Mint News App to get Daily Market Updates & Live Business News.