Transcription
Okay, so the Kim team just released K2.5, which is their first multimodal model. They're calling it the most powerful open-source model to date. If you go by the benchmarks, there is some truth to it. It beats giants like GBT 5, Gemini 3, and even Opus on a number of key benchmarks. And their main pitch is that it provides similar or better performance compared to some of the Frontier models at a fraction of the cost.
Now from what they have shared, it does seem to be a pretty strong coding model. Specifically fine-tuned front-end design, but this goes beyond a single model. They actually released a whole new system called agent swam. We have up to 100 different sub aents executing tasks in parallel. Now, this is not the first time Kimmy presented this multi- aent system called OK Computer, which can basically do tasks on behalf of the user. But now this system is going to be powered by K 2.5.
Now beyond Kim, this is going to be a busy end of the month for Chinese companies. Quen released Quen 3 Max thinking which is not an open-source model but it's state-of-the-art in this category. There's a new model from Deepseek called Deepseek OCR2 which just dropped and I think these companies are trying to push out a number of different releases before the lunar year.
Okay, we're going to do some tests later in the video to look at the actual capabilities of the model, but have let's have a quick look at some of the benchmarks. So, this is the first openweight model that crosses 50 score on the full humanities last exam, which is incredible because on the same benchmark, GP Gemini 3 Pro with high thinking level is getting 45.8. Now for on the agentic use cases like browser comp, it surpasses all the other models, even the proprietary state-of-the-art models, you could say there might be some benchmaxing going on. We'll see. But specifically, it seems like Kimmy is focusing on agentic use cases. Again, it's a very strong model. So on the Sweetbench verified we're getting around 76% 77% very close to state-of-the-art out there.
This model has multimodal capabilities and this comes in the form of both image understanding as well as video understanding and in both it has very close to Gemini 3 Pro. If this holds true, this is going to be an incredible feat because the pricing of this model is pretty amazing. Now one of their main focus with this is front-end capabilities. Gemini 3 Pro is probably one of the best model which can do front-end development out of the box and they are specifically focusing on front-end design. Some of the examples that they have shown here are actually pretty incredible. It doesn't look like a simple AI slop like we have been seeing. Here is another output. Again, it's pretty creative in terms of what it's capable of.
One thing which I was actually surprised that they are sharing some of their own internal benchmarks. All these frontier labs have their own internal benchmarks and usually they don't really report the results on these proprietary benchmarks but the Kimmy team is reporting those which is pretty good to see along with the model. I think the most important release from this is agent swam and this shows that scaling out is critical. You can't just scale up. They give the Kimmy model the ability to orchestrate multiple different parallel sub aents to accomplish a task.
Now it's just not through a agent orchestrator, but they actually trained it with parallel agent reinforcement learning where it learns to spin up to 100 different sub aents executing parallel works across up to 1,500 coordinated steps. And with this they're showing that using agent swam you can get much better results compared to the base Kim K2 version tools like clot code or Kim code could potentially use this architecture where the agent itself is trained through reinforcement learning on using these parallel agents. In general, a harness around an agent from the agent or the model creator is more important because they actually have done the reinforcement learning for those specific tools that the agent is supposed to be using. In here they show this very interesting result that using parallel sub aents not only can solve complex tasks but it can help you reduce the execution time as well.
Now if you see here the time taken by a single agent to solve complex task increases as the complexity of the task increases. But if you're using parallel sub aents that task complexity versus the execution time increase is pretty minimal. And this would make sense especially if you have very well definfined and welldesigned context management of these parallel agents where every agent has its own context window. This will not only enable you to solve complex tasks but also substantially reduce the number of tokens used by the overall system.
Now some details about the model architecture before we look at some demos. It's still a 1 trillion parameter model. It's an MO or a mixture of expert. There are 384 experts and at a time these experts are using only 32 billion parameters. The context length for this model is 256,000 tokens. I think this is more than enough for general programming tasks. These models are getting bigger and bigger. So, it's not for people like us to run them in our own garages. These are potentially for companies who want to use openweight models. Unfortunately, Kimmy is probably the only company that hasn't released a smaller openweight model yet, but at this point, they should.
If you want to test this model out, it's available on their website. You can use it for free. Right now in the model list, I can see Kimmy 2.5. There is an instant version, thinking version, agent, and agent swarm. I think this one is available for paid customers only. So, we're going to just quickly test the 2.5 on a few prompts.
Okay, so this first one is probably the most difficult one. I haven't seen any model except Gemini 3 Pro able to create this. So the prompt is create an animation of a crowd of people walking to form hello world I'm Kimmy as camera changes to bird I and we're specifically asking it to use 3.js.
Okay. So the thinking itself is very well structured. I think it's shows you the raw chain of thought. It says the user wants a 3JS animation. It talks about what exactly the requirements are. It thinks about what exactly it's supposed to create. Now it is coming up with a plan. This is very well structured. Now technical details. It is thinking about the structure of the code itself which is pretty good. And now it started writing code.
Okay. Now this is impressive. So I wasn't actually expecting this at all. Now it created this scene. I can't rotate it which is a bad thing. But something actually interesting happened. If you look at at the end it seems like it got it. Now, the main problem is that these words seems to be upside down. It seems like it got the kimit right though. And to be honest, I wasn't expecting this at all. But here's the thing. It's a multimodal model. So, we can just copy this. It seems like some of the words are upside down. I have attached an image. Also, make sure that you use a high contrast between the foreground and background. The people and the word formation is visible. Also there is no controls for rotating the images or changing the view. Add those as well.
Okay. So this is incredible. It is going through this and it's chain of thought. Seems like it has interled tool usage capabilities. So it's thinking about what exactly is wrong and also making use of the image as well that we provided. So it is able to understand image. Okay. This is actually a really pleasant surprise. This time it did add the ability to rotate it around and you can actually make out the words now. Seems like it fixed that mirroring issue. It did put the people in the correct order which is pretty incredible. I was not expecting this from an open weight model this soon.
Here's another prompt. This is the fox pod garden. It is fully functional. Although the UI uh design could definitely use some work but when it comes to prompt following it did a really good job.
Okay. Now one of the focus is front- end design. So I asked it to create a website. I wanted a new brutalist design and it went ahead and wrote about thousands line of code. Now here's what it actually created.
Okay. So the dark theme works which is pretty neat. also added these little animations which is pretty awesome. Everything seems to be functional. It added the ability to run code but this does not seem to be working. Overall I think it's probably one of the best output that I have seen from an open weight model when it comes to design but it's nowhere close to something like Gemini 3. I would rank it somewhere around Gemini 3 flash which still is really great.
Now, overall, I think it's a very solid release and sets the tempo for the rest of the 2026, both for openweight models as well as the proprietary close source models. Have a play with it. I think you're going to like the vibes of this model. Anyways, I hope you found this video useful. Thanks for watching and as always, see you in the next.