중국 오픈소스AI 또 공습… 성능까지 다 압도, 미국 빅테크 위기인가, 우리나라는? | 1조 파라미터 Kimi K2 압도한 알리바바 Qwen3 최신 업데이트, 중국발 AI 대전쟁

0:00 Yes, nice to meet you all. This is an error. A while ago, Kimi K2 from Moonshot AI, which has one parameter, was released as open source, so I dropped my horde value once again. DeepK, I was talking about how it's coming out again this time, and it's China's Alibaba Group. The Kuanyo series is actually an open source model that is widely used in our country. Since our country no longer has its own model, it often builds by bringing in a lot of open source, and as far as I know, many developer groups are using this to build various services . This is the third version of Kuen , Kuen 3, which came out before, and an updated version called version 257 came out, which has less than 1 parameter of Kimi K2, but its

0:45 performance is far superior to Kimi K2 . There are many parts where the performance is better than that of Claude 4 or various HPTs, so I 'm thinking that the Chinese AI is really scary, so I'm going to organize it today. In fact, whenever I deal with Chinese content, people who really don't like China come and the comment atmosphere becomes a bit hostile, and that's why I've been doing this, but in order to follow tech trends, we have to acknowledge and look at the areas where China is ahead . Just because you badmouth China does n't mean its technology is inferior. Oh, so I'm not saying I support China or anything like that. Let's look at these facts objectively, and compare the situation, China's strategy, and the strategies of American big tech companies . And at the end of the video, it is true that

1:31 open source models that focus on China are continuously emerging, and this is changing the landscape of the AI market . And most companies in our country are also benefiting. In fact, let's take a look at it from that perspective, in terms of how it opens a new chapter in the AI leadership race. Well, in the case of Kimikate, we put it on right away last week. It hasn't been that long. In doing so, it is also meaningful to open source such a huge model with parameters exceeding 1 page. Of course, bigger is not necessarily better . But since such a huge model has been released as open source, we can practice building this infrastructure by testing this huge model . Of course, this model may not be useful to everyone, but it can be a model that can be studied in its own right, which is why the

2:16 impact of open source is so great . Because none of the big tech companies open-source such a huge model . This is actually what China is aiming for. So, in Moonshot AI, we take Kimi K2 and release the parameters on this side, so we only get 1, but the performance is better than that . While doing so, they released an updated version of Quan 3 . It is said that Alibaba Group has been studying this organization called Kuan for a long time. It's not a model that suddenly appeared, it's something that AI research has been focused on for about 10 years . This was actually first released in April of the year before last under the name Tongyiqian N1, and it's been 23 years, so it's the same product. We

3:02 started building this ecosystem in earnest in August when we released 7 billion models as open source. But since this is open source and released, there are a lot of models . The basic models of Kuan are 1.5, 2, 2.5 and now 3 . Next, there are some that are released separately from Kuan Audio as Seven Billion, and as you can see here, there is a version exclusive to Kuan Mass Mass, as well as a version exclusive to Vision, and there are various specialized versions that are released through coding . This means that the Coen family has been continuously building specialized things in various modalities. In fact, when you think of Alibaba, aren't you talking about big tech companies along with other Chinese dime-a-dozen companies ? So, what I'm saying is that I'm putting all the research I've done on AI for a long time into this . So, if you look at the family of Coen 3 that came out first, it starts from 0.6 billion, and the most is 30 billion, and

3:47 235 billion, that is, 235 billion. There are so many. As you can see, these are models that have been distilled down to a very danceable size and made very small. These are for use as on-board AR or locally. You can see the target here, right? From cloud to something a little smaller for servers. Next, if you look at this, you can tell that it's a cell phone because it has notes on it. Then I said that MOE, that is, Mixture of Export, these days all the big models are going as ME models . Because the more you grow a model, the smarter it becomes according to the laws of scaling . That's why the Kimi K2 also increased the parameters to one side. Of course, it depends on how well and efficiently you do it, but in any case, this scaling needs to be done significantly, and since MOE is a combination of experts,

4:33 when you actually use it for a presentation, the parts that are activated are limited. Here too, the total model is 235 billion, but in reality, only 22 billion are activated when the model is released. Since only about 10 are activated, a much smaller portion can be used when generating output, which means that GPU utilization is also slightly lowered. Of course, you should use all of this when learning. At the same time, the overall knowledge capacity of the model can be maintained at its maximum, and the cost of publication, etc., can be reduced dramatically if only 22 billion are activated . So, you can see it as a version that takes practicality into account, and these days, not only Ku, but most models are approaching it this way . Even in the case of the meta's llama, it actively has Mixter of Export right now, so its context was 131K at the beginning.

5:19 Actually, 131K is a lot too. For those of you who don't know, since there are so many contexts, a context is a number that can be entered at once. So, it's said to be around 131,000 tokens, which is enough to hold several books . But this was confirmed to be 256K in the recent update, that is, the 2507 update that was announced recently . Because Kimi K2 said it supports that much, we expanded it and rescued it. In the case of Claude 4 Opus, the length of the connector has increased, so we have responded to this and matched it here. Actually, if you look at the contents of the hugging face, it says that you are in thinking mode. What this is, there's a model like OpenAI's 3 O4. Thinking models. You know that inference model, commonly called a reasoning model? Inference and reasoning are

6:05 both translated as inference in Korea, so it might be confusing. But you can think of reasoning as a model for thinking . I wish they would change this terminology sooner. I 'm writing about Infrance, but I don't necessarily think about it for a long time or think about it over and over again, but there is GPT 45 or version 4.1. You can think of it as a model that thinks in a single way. While talking about the 256K long context understanding yogurt, he said that it has been updated a lot across the board . It has increased further. When updating, there are some parts that actually require support from the GPU. You said that you updated various parts while updating the units that can be processed in one room. Let's take a look at this. Oh, isn't that amazing? You

6:51 compared it based on the Think King model. First of all, who is in last place? Blue. Blue is when Coen 3 first came out. This is a non-kicking model that has been slightly deteriorating in the past, but the Kimi K2 Moonshot AI, which was released last week, has a very high score of 75.1 . Even for something like CloudOffice 4, it's almost at this level, but CloudOffice 4 isn't open source. Kimi K2 is open source. But this time, the one I put out got a higher score than that . And an AI 25 score that is much higher than the Kimi K2 or Claude Opus 4 . So, we just used the existing MOE structure as is. We are using it as is, but we can see that the long context processing capability has been tuned more precisely, and the scores for complex benchmarks such as AI have increased significantly. Even in cases like this, it has

7:38 increased a lot compared to the existing Cohen, and things like Kimikate, Claude Offs, and Thinking Mode Deepseak, in fact, Deepseak was a hot topic, but that was because it was open source. But the research history of Alibaba is much older. I've been using Kuen very well, so I actually use Deepk a lot, but I use Kuen even more. That's the case in our country , actually. So Arena Hard 2, Arena Hard 2, and so on, are benchmarks related to user satisfaction. So, you can see that in math, user satisfaction, and things like that, it's almost ahead of all the scores. Of course, it's not like we're ahead of everything. I've shown you the ones above, but in the case of MMLU Pro Redux, there are also some that are better than Claude Offs 4 Nuncking. Oh, but the thing is, most of the rest are things like widespread distribution and reasoning, and so

8:25 on. Isn't it amazing? Of course, there are also things that Kimikate is ahead of. But this time, there have been a lot of improvements made to the agent side in relation to Kuan . Can you see that there are a lot of updates starting from the one on the left? What you need to be aware of here is that Korean is also very good at multilingualism across most benchmarks. This is also organized in the Korean leader report, but in fact, the highest of all of the above is currently Perplexity Sonar Pro. This is currently being evaluated as the best Korean language . Of course, it's based on a specific test score, but other than that, everything else is like Entropy Cloud Offset 4 is better in Korean, Gemini 2.5%, GPT 4.5, Open AI 51 won, etc. Most of these are currently being serviced like this, but if you zoom in a bit on the upper part, you can see that Grock 3 is also listed here, and

9:12 Grock 4 has also come up to here. Purple is really nice. So I think the satisfaction level of Purpleless is particularly high in our country . Now, do you see this? Quen 3 235 Billion 2507 This is the updated one, can you see it performs better than Deep V3024 ? This has also come up a lot . Of course, there are also some things that make Kimi K2 a little better in Korean. Among the models in our country, SKT's AX4.0 0 is said to be quite good, and Naver's open source models are probably the ones below because they only have small models, and Kakao does n't have any large models either. LG is here with the X41. Then KT KT will be down there and upstage. Upstage is still here. You could say that Upstage and Coen 3, which are now open source, are at almost the same level . So here is how to install it, and I'm

10:00 thinking of installing a small model and running it locally, but from that perspective, this Kimi K2 is actually the deep-seek of yesterday's J2, isn't it? Because I did it with a really huge model, and it was updated and disappeared like this, so I should be less excited now. It's not easy for you to buy it by referring to YouTube. Oh, I'm sorry for being so rude. I will try to do this a little more neatly so that it doesn't bother you all. Please watch it for fun. But the point is this. It's 235 billion less than the parameter on page 1, so it 's less than 1. The fact that performance can be good even with numbers less than 1 is not something that can be achieved simply by increasing the parameters. This also

10:45 shows that it is very important to be good at FineTwin and to continuously learn data. This has a big impact on everything from AI services to open source, and especially in the case of Coen, DeepSeak started this in the open source camp, but Coen is way ahead. There are probably some areas where we can judge that various models are being distributed well . In other words, when we talk about China's AI, there are some areas where we need to focus more on Kuen than other areas . Alibaba is building a huge number of servers and investing a huge amount . When I see that, I think that China's rise in AI, that is, when it comes to going up against big tech, the offensive of open source is really scary . First of all, the government is providing a lot of financial support and actively promoting AI at the national level. The scale of it is truly enormous. In fact, the United States is also pushing it very hard. Well, they're investing a lot in Star and all, but

11:32 of course, because of the US's technology sanctions, they can't import expensive GPUs in large quantities anymore . Of course, they say that H20 is now useful, but even then, Nvidia will just sell the H20 it currently has. There is talk like this. Oh, of course, our country's memory company, that is, H20 is known to have Samsung Electronics' HBM3 in it, and there is a new model called H2. I think reporters are just calling it H2 these days, but it also includes HBM3e. I understand that SK Hynix is putting that in now. So, despite the restrictions imposed by the US, if they invest so heavily in open source, many people in our AI community will build systems based on this. In fact, there are many small

12:17 startups and startups that are not related to AI . Startups don't just do AI. You can do general office work, commerce, and other things, and to increase productivity, you can utilize this open source model a lot . In fact, in order to obtain knowledge specific to a certain company, we do fine tuning based on domain knowledge and pull in lag, and so this ecosystem is created based on this. Then, we will have no choice but to gradually become dependent on China's open source . That's why, in our country too, there are stories like "Sorin AS" and "Sorin A." So it is difficult for us to actually invest that much. Actually, it would be difficult. If that kind of capital investment is difficult, then in order to secure that kind of technology, rather than focusing on this kind of legal model, as we saw before, SKT, Saltlux, and Upstage,

13:02 these models are also working hard and our researchers are working hard. I also have this thought that we need to find some direction in developing AI models by targeting specialized areas such as medicine, law, and semiconductors. But there are many data sovereignty issues, not just AI models. These days, even Korean companies are fighting among themselves over this, so it is important to secure quality Korean language data by industry. So, we need to create a model that is suitable for our environment and can be well-established in our country, but we may be fighting over the definition of AI. Because each company has a different understanding of this. But rather than saying that is important, I think it is important to have a form in which we can actually provide services and do business in our country . So, in the end, the

13:50 emergence of AIs like Wen 3 and Kimicate is not simply about AI being smart and just emitting the right radiation, but rather AI itself is becoming a tool that can enter certain key industries and even protect national security. If it becomes so pervasive by being open source, there is a risk that the company will become dependent on it at the moment when people say, “Our company can’t run without it.” So, these technological changes are coming now, and it seems clear that we need to think about how we should respond in our country and how we can secure technological sovereignty. But from that perspective, let me summarize the news I got today . Andae Gong's Instagram account is unyeoltaek. We deliver short news like this quickly, so please receive it often.