Deepseek R2 & V4 NEW Leaks: 98% Better? 🤯

There's been a bunch of new leaks about the new Deep Seek R2 that is coming out. You can see the details right here. This is a big shift away from the US supply chain if it is confirmed as true. So, you can see all the information right here where it came from etc. And the unit costs and the actual cost of Deep Sea Gatu in these leaks according to them is going to be 97% cheaper. So if we break this down, really good tweet by Bindu. Basically the rumors are that number one R2 is a 1. 2 trillion parameter model, double the size of R1. The cost will be a fraction of GPT4 row and it's trained on 5. 2 PB of data. And the other craziest thing is that they have overcome not having access to Nvidia chips, which means that it's not relying on, for example, like the US to actually launch this. So, here's the full details. I'm going to

break it all down for you. This was last updated April the 28th. And basically, recently, as you can see in this image right here, a wave of online leaks and rumor posts claimed that Deepseek is going to unveil its R2 new language model. So, let's cover what the leaks actually claim, the evidence, and expert opinions, the question, the legitimacy of it, a timeline of the events, and insight into what we can realistically expect from Deepseek in the future. So here's the full details of the leak. So 1. 2 trillion parameters with a hybrid mixture of experts model 3. 0 architecture and also up to 97. 3% cheaper than GPT4 Turbo. So that's 0. 07 per million input tokens compared to GPT4 Turbo which is 0. 27 per million output tokens. It's based on 5. 2 2 pabytes of training data with 89. 7% on the C eval 2. 0 suite. The interesting thing here as well is the hardware used. So, it's trained on Huawei's Ascend 910B accelerators which bypass Nvidia and it also has an 83% size reduction. Now, if you're wondering what that means in English, basically it could make full scale enterprise LLM deployment 10 to 100 times cheaper than existing models. It's also positioning China as an AI sovereign independent of US chip supply chains for Nvidia and it sets a new bar for reasoning and multimodal applications. Couple of the sources used here. So tech power up they've covered the full leaks and also

moniker. im. However, this is what the YouTube videos won't tell you. So there's a lot of evidence pointing to these leaks as fake news. For example, what people are saying on Reddit, on Twitter, etc. is highly likely to be fabricated. And the reasons for this are number one, there's a lack of credible sourcing, right? These leaks originate from Chinese and finance stock trading forums, not like verified insiders or proven technical sources. The other thing here is that some people are saying this is conspiracy theory of stock market manipulation suspicion and several posts explicitly reference or promote concept stocks raising red flags about the intent, right? So, you got to be careful about that, too. And the other thing here as well is the timeline doesn't really add up. So there's insufficient time for a 1. 2 trillion mix architecture. The training data for Deep Sea Car 2 claims 5. 2 PB which is way higher than most leading labs. That's unfeasible too. And the benchmark claims right the cited numbers for the benchmark claims way exceed like state-of-the-art models which is quite interesting as well. And the other thing to note here is that Deepseek's own public accounts have denied R2's launches, labeling the stuff as fake news. So you got to be careful who you follow and what you believe, etc. There's also a few big names in the industry who have expressed strong skepticism and outright disbelief. So this is a leak translated into English from Chinese. You can see the information right here. And if you're wondering, okay, what's the timeline of the leaks and responses? So this

basically came out around about 25th of April. And the leak appeared on Chinese finance forums and then was translated. Loads of AI news channels and YouTubers started covering the leak. And there was an earlier wave of Deep Seek denying the R2 release date rumors. So just something to be aware of here is like there's no public release yet. There's no official technical paper. There's no benchmarking from Deep Seek or Standard LM evaluations as of right now. So just something to be aware of. It's like you're going to see a lot of news about these Deep Seek R2 leaks. what they mean, etc. But so far, nothing has been confirmed. What I will say is

that Deep Sea Car 2 is 100% coming. They've already released an awesome model with Deep Sea Car 1. They're doing a lot of like hybrid models as well as Open Reason models. They probably will shift from Nvidia to Huawei or other Chinese suppliers for the chips, especially because of the current situation in the world right now. If you actually want to get like the live details etc you can just go to hugging face. So you can actually go to their official hugging face profile which you can see at Deepseek AI over here and then also you can go into GitHub and you can see the latest releases right here. For example, the latest releases so far have been Deepseek R1 and V3. So overall what does this mean? I would say the Deep Seek R2 leaks are fascinating and very well crafted but it's a bit of a viral rumor campaign right now. No one else is saying this on YouTube, but I just wanted to come out and tell you transparently. Probably fueled by a lot of people saying this is fueled by like speculative stock market activity in a dash of like copium about AI progress. But it's a reminder of the hype and the hope and the competition in global AI markets, especially when you compare the US versus China leaps. Right? If we actually look at the competitive benchmarks here and if you compare Deepcar 2 versus GT40 versus Claw 3, right, you can see them on a chart right here. And if the leaks are true, this is what it would look like in terms of parameters. It's just silly, right? And I can't see that being the case, especially because if you look at the history of Deep CKI and where it came out from, it's the acceleration of that would be insane, right? Okay. So, for example, they were only founded in like 2022 and then 2023 they released Deep Seek version 3 February and then all of a sudden we jump to that and the parameters versus some of the best models in the world. Doesn't really make sense. What I will say is that in the future, obviously Deep Seek R 2 is going to be coming out soon and Deep Seek R4 and Deep Seek version 4 will probably come out as well because we've already seen Deep Seek

V3. 1. So, thanks very much for watching. If you want to get access to a community of people focused on making money and saving time with AI, feel free to get the AI profit boardroom. This community is updated pretty much every single day with new SAP systems, templates, workflows, etc. You can see just for April alone, look at how many different systems we've added for Deep Sight, Gemini, Lindy, Rico, Gen Spark, etc. All the latest tools, models, SPs, and workflows are directly inside the AI profit boardroom and inside the community as well. You can ask any questions you have, if you ever get stuck, etc. You can see all the updates that we dropped this week alone. There's many more coming. We pretty much update it every single day. So, feel free to get that link in the comments description.

And if you want to get a free one-to-one SEO strategy session that shows you how we take websites from zero to 145,000 visits a month and generate hundreds of thousands of dollars in sales on autopilot, feel free to get that. We will give you a custom tailored game plan on the call based on what's working for our happy clients like you can see right here on how to get more free leads, more free traffic, more free customers directly from Google to boost your profit margins and to help you rank inside AI search engines as well. Feel free to get that link in the comments description. Appreciate you watching. Bye-bye.