00:00Leonardo Silva Reviewer
00:07All right, welcome, everybody, to LlamaCon.
00:23The groundswell of support for Llama has been awesome.
00:31We announced 10 weeks ago a billion downloads after the release of Llama 4.
00:35In just 10 weeks, that number is now 1.2, which is exciting.
00:43Today, we're announcing that you can now start using Llama with one line of code
00:47because we're releasing an API.
00:53I'm going to invite up Manahar Polari and Angela Fan, who are our very first Llama engineers,
00:59to walk you through how that works.
01:01Come on up, guys.
01:02Our goal from the beginning for Llama 4 API has been it should be the fastest and easiest
01:08way to build with Llama.
01:10Not only that, as you saw, the derivative models, it should be the best way to customize these
01:16models.
01:17And going one step further, you should be able to take these custom models with you wherever
01:22you want.
01:23No lock-in ever.
01:24The Llama API is really good at fine-tuning for your product use cases.
01:29And the best part, you have full agency over these custom models.
01:34You control them in a way that's not possible with other offerings.
01:38Whatever model you customize is yours to take wherever you want, not locked on our servers.
01:43This opens up a lot of possibilities for people that need to build and deploy custom models.
01:48It's very important.
01:49Thank you so much for joining us.
01:50We can't wait to see what you build with Llama 4 and the API.
01:54Back to Chris.
01:55This is part of how I think open source basically passes in quality.
02:07I mean, part of the open source vision is just this belief that there are going to be
02:10many, many different AIs.
02:11Right.
02:12Pick one big model and you use that and there wasn't that much fine-tuning and so on going
02:17on.
02:18With the reasoning time compute.
02:21So if another model, like DeepSeek is better or Quen is better at something, then now as
02:27developers, you have the ability to take the best parts of the intelligence from the different
02:32models and produce exactly what you need, which I think is going to be very powerful.
02:36So we have to automate it somehow.
02:38So it's coming up.
02:39I think evals and data sets are going to be super.
02:42Thanks, Mark.
02:43All right.
02:44Thank you all.
02:45Power Meta AI, which is used by almost a billion monthly users.
02:53And speaking of Meta AI, we announced an app this morning, which we launched.
03:05So we wanted to focus on pushing the limits and a fresh take on how people could use AI.
03:10We were very focused on the voice experience, the most natural possible interface.
03:15So we focused a lot on low latency, highly expressive voice.
03:19The experience is personalized, so you can connect your Facebook and Instagram accounts.
03:24And the assistant will have a rough idea of your interests based on your interaction
03:28history.
03:29You can remember things you tell it, like your kids' names and your wife's birthday and all
03:33the other things you want to make sure your assistant doesn't forget.
03:38We added an experimental mode, which is full duplex voice.
03:42This means it's trained on native dialogue between people in speech.
03:46So rather than the AI assistant reading a text response, you're actually getting native
03:50audio output.
03:51And full duplex means that the channels are open both ways so that you can hear interruptions
03:56and laughter and an actual dialogue just like a phone call.
04:00So it's pushing the limits what we think of what's possible with a natural voice experience.
04:08Thanks.
Comments