Meta releases Llama 3.1 open-source AI model to take on OpenAI

Back in April, Meta teased that it was working on a first for the AI industry: an open-source model with performance that matched the best private models from companies like OpenAI.

Today, that model has arrived. Meta is releasing Llama 3.1, the largest-ever open-source AI model, which the company claims outperforms GPT-4o and Anthropic’s Claude 3.5 Sonnet on several benchmarks. It’s also making the Llama-based Meta AI assistant available in more countries and languages while adding a feature that can generate images based on someone’s specific likeness. CEO Mark Zuckerberg now predicts that Meta AI will be the most widely used assistant by the end of this year, surpassing ChatGPT.

Llama 3.1 is significantly more complex than the smaller Llama 3 models that came out a few months ago. The largest version has 405 billion parameters and was trained with over 16,000 of Nvidia’s ultraexpensive H100 GPUs. Meta isn’t disclosing the cost of developing Llama 3.1, but based on the cost of the Nvidia chips alone, it’s safe to guess it was hundreds of millions of dollars.

So, given the cost, why is Meta continuing to give away Llama with a license that only requires approval from companies with hundreds of millions of users? In a letter published on Meta’s company blog, Zuckerberg argues that open-source AI models will overtake — and are already improving faster than — proprietary models, similar to how Linux became the open-source operating system that powers most phones, servers, and gadgets today.

He compares Meta’s investment in open-source AI to its earlier Open Compute Project, which he says saved the company “billions” by having outside companies like HP help improve and standardize Meta’s data center designs as it was building out its own capacity. Looking ahead, he expects the same dynamic to play out with AI, writing, “I believe the Llama 3.1 release will be an inflection point in the industry where most developers begin to primarily use open source.”

To help get Llama 3.1 out into the world, Meta is working with more than two dozen companies, including Microsoft, Amazon, Google, Nvidia, and Databricks, to help developers deploy their own versions. Meta claims that Llama 3.1 costs roughly half that of OpenAI’s GPT-4o to run in production. It’s releasing the model weights so that companies can train it on custom data and tune it to their liking.

Unsurprisingly, Meta isn’t saying much about the data it used to train Llama 3.1. The people who work at AI companies say they don’t disclose this information because it’s a trade secret, while critics say it’s a tactic to delay the inevitable onslaught of copyright lawsuits that are coming.

What Meta will say is that it used synthetic data, or data generated by a model rather than humans, to have the 405-billion parameter version of Llama 3.1 improve the smaller 70 billion and 8 billion versions. Ahmad Al-Dahle, Meta’s VP of generative AI, predicts that Llama 3.1 will be popular with developers as “a teacher for smaller models that are then deployed” in a “more cost effective way.”

When I ask if Meta agrees with the growing consensus that the industry is running out of quality training data for models, Al-Dahle suggests there is a ceiling coming, though it may be farther out than some think. “We definitely think we have a few more [training] runs,” he says. “But it’s difficult to say.”

For the first time, Meta’s red teaming (or adversarial testing) of Llama 3.1 included looking for potential cybersecurity and biochemical use cases. Another reason to test the model more strenuously is what Meta is describing as emerging “agentic” behaviors.

For example, Al-Dahle tells me that Llama 3.1 is capable of integrating with a search engine API to “retrieve information from the internet based on a complex query and call multiple tools in succession in order to complete your tasks.” Another example he gives is asking the model to plot the number of homes sold in the United States over the last five years. “It can retrieve the [web] search for you and generate the Python code and execute it.”

Meta’s own implementation of Llama is its AI assistant, which is positioned as a general-purpose chatbot like ChatGPT and can be found in just about every part of Instagram, Facebook, and WhatsApp. Starting this week, Llama 3.1 will be first accessible through WhatsApp and the Meta AI website in the US, followed by Instagram and Facebook in the coming weeks. It’s being updated to support new languages as well, including French, German, Hindi, Italian, and Spanish.

While Llama 3.1’s most advanced 405-billion parameter model is free to use in Meta AI, the assistant will switch you to the more scaled-back 70-billion model after surpassing an unspecified number of prompts in a given week. This suggests the 405-billion model is too expensive for Meta to run at full scale. Spokesperson Jon Carvill tells me the company will provide more information on the prompt threshold after it assesses early usage.

A new “Imagine Me” feature in Meta AI scans your face through your phone’s camera to then let you insert your likeness into images it generates. By capturing your likeness this way and not through the photos in your profile, Meta is hopefully avoiding the creation of a deepfake machine. The company sees demand for people wanting to create more kinds of AI media and share it to their feeds, even if that means blurring the line between what is discernibly real and not.

Meta AI is also coming to the Quest headset in the coming weeks, replacing its voice command interface. Like its implementation in the Meta Ray-Ban glasses, you’ll be able to use Meta AI on the Quest to identify and learn about what you’re looking at while in the headset’s passthrough mode that shows the real world through the display.

Aside from Zuckerberg’s prediction that Meta AI will be the most-used chatbot by the end of this year (ChatGPT has over 100 million users), Meta has yet to share any usage numbers for its assistant. “I think the entire industry is still early on its path towards product market fit,” Al-Dahle says. Even with how overhyped AI can already feel, it’s clear that Meta and other players think the race is just beginning.