META 的逻辑是高速增长、宽阔护城河与低估值，而凭借 Muse Wonder 前沿模型，它如今在 AI 公司中处于领先地位，并拥有唯一的个性化智能体 Muse。它仍被低估，应以 26 至 27 倍远期市盈率交易，预计估值还将进一步扩张。

Mark Zuckerberg ordered a watermelon. He teased that this morning. In reality, that was the code name for the new model that he was going to release. The new model's called Muse Wonder. The stock is now up 4% on the day.

Now, moving on, we get into some Meta news. Meta continues trading with a lot of confidence. This stock has a lot of momentum. It's moving up 3 to 4% on the day after Mark Zuckerberg said on a on Twitter.

He posted that Muse ordered me watermelon. How thoughtful with a picture of a watermelon.

Alexander Wing has said internally that Meta is finally catching up to OpenAI. They're finally catching up to the Frontier models. And how are they catching up to OpenAI? Well, with this model, which is Watermelon, that's the code name for it, but its official name is Muse Wonder.

Mark Zuckerberg released today, introducing Muse Wonder, our new family of Frontier models with state-of-the-art performance for personal and professional agentic work.

To put it simply, Muse Wonder is super good and Meta has been leaprogging ahead of the competition at a rapid pace. You can see a visual here. You have Muse Spark, which was behind the pack.

Then we have Musepark 1.1, then 1.2 two that leaped ahead two positions. Then we have Musepark 1.3 that leaped ahead another three positions. Then we have Muse Wonder. Then we have Muse Wonder full which is now literally ahead of all the models except for Claude Opus 5.5 and it is barely behind Claude Opus.

So you can see how fast this is gaining on the other models. Now it's virtually at the top of the pack. It's right there with Anthropic's best current model.

Meta also addressed concerns of the benchmarks because we know a lot of models can be trained specifically to perform well during some like benchmark test. This is benchmaxing is what some of them are calling it.

Now the important thing here is that both of the models improve similarly on outofdistribution tasks as we scale test time compute but given its distribution performance muse wonderful full performs even better on unfamiliar tasks in Muse Spark 1.3 uh trend would predict.

So this is like a weird way of saying that they're not trying to game the benchmark. They're not benchmaxing by just making their models good at a couple tasks that are particularly curated to advance in a benchmark.

No, even when you throw all different random tasks, things that it's unfamiliar with, it also has incredibly good performance across the board.

Now, they give examples of this model like planning a run. You can see it just open up someone's calendar. It makes a new calendar entry and then you can see the the things like going through all the code and iteration looking at a map of of the location and then it literally draws out a run depending on the person's desired distance they want to run.

It it puts all the information into the notes. it has the time for it and everything based on what they want. So if you want Muse to map out a run, you can do that. So this is like one example of I think how Mark Zuckerberg and Alexander and all these people envision people actually using Muse.

It's just to do the kind of annoying things that you don't want to do for you, like planning a run. You probably want to go on a run, but you don't want to actually go and figure out where to run and the distance and where would be a good run that's different from the last one you did.

So you could have it do that kind of annoying part while you get to focus on the fun part which is going out and actually going on a run. I think that's how they envision it.

Of course there's going to be a lot of people that try to use it in bad ways. There's a lot of controls are doing to make sure that the models are safe that they ask for permission before taking tasks that they don't display your information or your personal stuff anywhere. uh there's an enormous amount of effort doing that because if Meta loses trust with their models, if Muse isn't trustworthy, then people aren't going to use the Muse app and so far that's one of the best things they have going for them.

I see overall that this is another huge step in meta. I become more and more bullish on this company over time. I think the base thesis for the company is that it was growing really fast.

It has a wide moat and it's at a cheap valuation. But now you have a company that also is leading in the pack of AI companies. It has a truly a truly frontier model, one that is right there with Claude.

And it's the only company that has a personalized agent, which is Muse. Nobody really looks at ChateBT or Claude as being their personalized agent. They use it for different business applications, maybe answering questions, but Muse has become the deacto agent.

So, I believe that Meta still has a significant opportunity here.

When I look at this company, it's one of my most exciting investments today. I think the company should be trading at a 26 to 27 Ford PE ratio. I think there should be more multiple expansion. I still believe Meta is undervalued.