DeepSeek V4 Released - Best AI Model Ever

追加: 2026-05-05

[00:00:00]Finally, guys, Deep Seek V 4 Pro and Flash are finally out. 1.6 trillion parameters on Deep Seek V 4 Pro. It's competitive with the world's best model, GPT 5.4, Opus 4.6. And you can see, for example, Sweet Benchmark Terminal Bench, it's very competitive with the Frontier.

[00:00:20]This is absolutely insane, and it's trained on very nerfed Nvidia GPUs and Huawei GPUs. If you want to learn how to do AI research and build large language models like Deep Seek V 4 from scratch, join our school. Uh link below the video. And there is community that will help you become AI researcher.

[00:00:38]It has a very rich world knowledge. We all know that Gemini 3.5 is very good at just like knowing so many facts about the world. So, Deep Seek V 4 is actually uh only uh trailing behind Gemini, which is so crazy that an open-source model is capable of doing this. In China, uh Deep Seek is considered the best company by uh by other companies as well. Every other company in China thinks that Deep Seek is the is better than them.

[00:01:08]These benchmarks are absolutely insane.

[00:01:10]It's the best open-source model, even if Kimmy uh 2.6 just got released, Deep Seek still beats it.

[00:01:16]And I also believe, guys, that uh Deep Seek is working closely with Chinese government to advance Chinese chips.

[00:01:23]This is why it took them so long to release the next model. They are probably working with Huawei and all of the other chip makers in China to make Chinese chips very competitive.

[00:01:33]And I think Deep Seek V 4 is heavily trained utilizing uh Chinese chips. And I also think Huawei and chip makers are working closely with Deep Seek to uh improve their chips for Deep Seek training, for LLM training, etc. So, that's why it's taking them so long. But I think they made big improvement now.

[00:01:51]So, I think this is the uh biggest open-source model that we have by now, and this performance is absolutely insane. It's so good. Look, it's competing with the best models in the world right now.

[00:02:03]Look at the computing improvement they made. Previously, Deep Seek V 3.2 uh had this scaling, this line, but look, as you add more tokens, more context, you see how it's scaling a lot less. You need a lot less compute now with Deep Seek V 4 because it's using DSA and token-wise uh compression.

[00:02:24]And 1 million uh token context is now standard. I think we're going to start using Deep Seek V 4 for agentic coding, for everything. It's going to be so much cheaper, and it's going to create a lot of pressure on GPT, on Anthropic. I'm curious to see what Anthropic will do now because Anthropic is pulling away all of the uh coding like Claude code and stuff. So, we'll see how what effect this has.

[00:02:49]And it can seamlessly integrate with uh Claude code, open Claude, open code, etc. There is an absolutely huge technical report here going on.

[00:03:00]So, they're using manifold constrained hyperconnections. That's uh one of their papers.

[00:03:05]They're also using some new architectures. I'm compressed sparse attention, heavily compressed attention.

[00:03:10]I'm I didn't see this before.

[00:03:13]They're also using multi-token prediction, same as V 3. They're also using compressed sparse attention. This is the architecture. Uh maybe we will make more videos on these.

[00:03:23]This paper is an absolute goldmine of knowledge. There is so much stuff here.

[00:03:29]All of the math theory on creating LLMs.

[00:03:32]I guess this video is more like uh news introduction. So, I'm going to make more videos about technical deep dives, about research that can be done from here. I also do research myself, by the way, I'm researcher. So, uh just subscribe and stay tuned. I'm going to uh start making more like math, technical, coding videos on all of this, explaining all of this.

[00:03:51]You can join our school if you want to learn all of this math, if you want to learn how to do AI research, build LLMs from scratch. And there is community, link below the video. Start your free trial right now. I'm always here answering, replying, adding uh notes, questions, videos, tutorials, etc. So, see you in the community.