Is Enthusiasm For Meta’s Llama Slowing Down?
To compete with OpenAI and other artificial intelligence developers, Meta Platforms CEO Mark Zuckerberg needs cloud providers like Amazon Web Services to convince businesses to use Meta’s AI model Llama.
So far, that has been a mixed bag, as my colleagues reported on Wednesday. And we just noticed another potentially troubling sign about Llama’s popularity.
Llama 3.1, the latest version of the flagship model, has been downloaded a total of about 3.6 million times from Hugging Face in the first month or so since its release on July 23, according to data on Hugging Face’s website. That’s far fewer than the 5.8 million downloads Llama 3, an older and less advanced model, had in the first month after its release in April.
Developers may still be figuring out how they want to use Llama 3.1, but one would expect more enthusiasm for a better model, not less. By contrast, here’s a slightly different metric: The newer versions of Mistral's smallest open-source model, Mistral-7B, has been downloaded more times in the last month than the first version of the model has.
What’s going on?
It’s hard to tell, but rivals such as OpenAI and Anthropic continue to push out improvements to their models as well. And unlike with Llama, which is open-source (or open-weight, depending on which terminology you prefer), lots of developers are paying to use the closed-source options. Despite being free, using Llama can sometimes cost more than using OpenAI’s models, as we previously reported.
One problem for Meta is that the top two cloud providers, Amazon and Microsoft, have special arrangements with Anthropic and OpenAI, respectively, to share revenue from model customers. (The cloud providers also hold big equity stakes in those startups and use their technology in their products.) And Google, the third biggest cloud provider, is probably incentivized to sell developers on its own model, Gemini.
A spokesperson for Meta declined to comment on the apparent slower downloads of Llama 3.1. But the trend could speak to why Meta is publicly and privately pushing AWS and other enterprise technology providers to promote Llama.
Zuckerberg has said that Meta has previously taken a cut of the revenue cloud providers generate from customers running Llama on cloud servers and that “it makes sense that we share the upside of that somehow.”
We don’t know if those revenue sharing arrangements have continued. But something tells us that if Zuckerberg wants any chance of changing the pecking order of AI models that cloud provider salespeople promote to customers, he will probably have to forgo that cash. That way, the cloud provider would keep all of the revenue from servers it rents out to users of Llama.
The Meta spokesperson declined to comment on this subject.
Meta itself doesn’t sell cloud computing services, and enterprise sales and marketing have never been a strength of the owner of Facebook, Instagram and WhatsApp. So it’s difficult to imagine Meta directly selling or providing Llama to developers.
But Zuckerberg has a dilemma on his hands. If he agrees not to make much, if any, money from enterprises using Llama, that may be a tough pill for investors to swallow. Meta is spending tens of billions of dollars developing Llama to aid its core social media apps and advertising business, and it needs to show a return on those investments. Zuckerberg has said that if more consumers and businesses embrace Llama, it will help Meta with that mission.
He also clearly wants to pull the rug out from under closed-source competitors so that they can’t make money from selling their AI models—in other words, to commoditize the models so that more money will be generated from applications that use AI models (like Meta’s) rather than from the sale of AI models themselves.
But making Llama as good as OpenAI’s models without a new funding source—such as sharing in revenue from cloud providers’ Llama sales—may be hard to do without also breaking Meta’s balance sheet.