Kimi K3 obté 62/80 a KingBench: és realment millor que Fable 5?

Hi, welcome to another video. So a few days ago, the Moonshot team reached out to me and gave me early access to their upcoming Kimmy K3 model. I have been testing it for the past one or two days now. And to be honest, this model is kind of crazy. It is easily one of the best models that I have used. And in this video I'll show you how it performs on my benchmark and share everything that I have found while testing it. Now, before we get into it a quick disclosure, Moonshot, the team behind Kimmy, is sponsoring this video, but they have just paid for the video slot. They haven't seen this video before, it's release, and they have no control over what

I say here. So everything you're about to hear is based on my own testing, and it's pretty accurate to my actual experience. Now, let's get right into it. So KimiK3 is the next model in the Kimi series from Moonshot. You might remember KimiK2 and K2.7, which were really good open models that a lot of people used as their daily drivers for coding. K3 is the successor to that, and from what I can tell after these two days, it's a massive jump. Now, one thing that I don't have any information on is whether this model is going to be open or not. They haven't given me much info related to that part, so I can't confirm anything right now. I'll pin a comment below once I have that info, so keep an eye on that.

Anyway, now let's come to the testing. As always, I ran it through my Kingbench questions. These include some front end and animation tasks, some three JS tasks, and SVG test, a math problem, a long horizon, agentic task, and one really hard 3D task at the end. Each question is scored out of 10, and at the end, we'll look at the final chart to see where it lands against the other models. There's also a new question as well. The first question is the elevator simulation. Basically, the model has to build a simulation where you can spawn people on different levels.

There are three elevators. Each elevator can only take one person, and the people left behind should catch the next one. It also needs tooltips showing each person's target floor when you hover over them. Let's send it and see, okay, so KimeK3 gets a nine out of 10 here. The simulation works, the elevator logic is correct. People wait properly and get picked up one by one, and the animations are smooth. Only some minor polish issues keep it from a perfect score. For reference, Opus 148 got a 10 on this one and Fable 5 got a 9.

So K3 is right up there with the top models. This is pretty amazing to be honest. The second question is the 3JS contact lens case. It should be a 3D model of a contact lens case with pronounced L and R caps. And you should be able to click on the caps to open them up. Let's send it and check. So here it gets a 5 out of 10. The model is there. The caps are labeled, But the opening interaction is a bit janky. To be fair, this is a test where almost every model struggles. Fable 5 is the only one that really nailed it with a 10, and Opus 4.8 got a 7, so a 5 here is still around the middle of the pack. The third question is the 3JS folding table. There's a slider, and as you drag it to the right, the table should unfold and to the left, it should

fold back. It needs to be seamlessly animated and properly 3D. Let's send it, and this one is really good. The folding animation is smooth, the geometry holds together, and the slider control works exactly as expected. This gets a 9 out of 10, which ties it with Fable 5, and G-Lm 5.2 at the top for this question, above Ocas 4.8, which got an 8. So this is kind of great. The fourth question is the SVG of a panda eating a burger. Let's send it and see. Okay, this is actually one of the better pandas I've gotten. The proportions are right.

The burger is recognizable, And it actually looks like the panda is eating it. It gets an 8 out of 10 which is the highest score on this question, along with grock 4.5 and GLM 5.2. Even Opus 4.8 only got a 6 here. So that's great. The fifth question is the Bowanaro Simulator game, four targets and hitting the targets in the lowest time should put the player on the leaderboard. Let's send it and check. The game works. The arrow physics field decent.

The leaderboard updates properly. It gets an 8 out of 10, Opus 4.8, got a 10 on this one, so it's slightly behind there, but it's still a very solid result. The sixth question is the hard math problem about counting permutations of ordered pairs, and the answer should be 2034060. This is a question that most models completely fail. GPT 5.5, Opus 4.7, and Gemini 3.5, flash all got zero on this. Let's send it and see. When KMEK3 gets it exactly right, 2460, full 10 out of 10. This puts it in the same club as Opus 4.8, Fable 5, and the GPT 5.6 models. This is quite great to see because it shows the reasoning is genuinely strong and not just

tuned for front-end stuff. The seventh question is the big, long horizon one. Generate a dataset of facts about pandas, finitune, a gemma2b model on that dataset, Then build a local web UI that generates a pan-to-fact on every refresh. All of it done autonomously from scratch. Let's send it and see. And this is honestly where KimiK3 shines the most. It generated the data set, set up the finituning, ran it, dealt with the errors that came up along the way, and delivered a working local web UI at the end, full 10 out of 10. And this matches my overall experience with the model.

It grabs a task by its head and just doesn't give up until it's finished. It doesn't stop halfway and ask you what to do next. just keeps going. The eighth and final question is the 3D wristwatch, a full-working wristwatch in 3D that shows the actual current time with smooth second, minute and hour hands, a date and day option, and two time zones. This is the hardest question on the bench and no model has done well on it. Let's send it. Chemike 3 gets a 3 out of 10. The watch renders and the hands move, but it's not the full-working watch that I asked for. For context, the best score on this question is Fable 5 with a 4 and models like GPT 5.5 mini max M3 and Mewsbar got a flat 0. So a 3 is actually one of the better attempts, believe it or not. So this is the final chart and Kami K3 ends up with 62 out of 80, which is 77.5%.

That puts it in third place overall, just behind Opus 4.8 at 80% and Fable 5 at 82.5% and clearly ahead of GLM 5.2 at 75% GPT 5.6. at 71.25% and grok 4.5 at 65%. But here's the thing. The benchmark numbers don't tell you the full story. In my actual day-to-day usage over these two days, this model feels way better than Opus 4.8 and GPT 5.6 sole. The benchmark is mostly single-shot tasks, and on those, the top models are all close, but the moment you give it a real, messy, long horizon task, K3 pulls ahead of both of them.

It's still a bit worse than Fable 5 overall. If you ask me, but for a Kimmy model, it is kind of crazy that we're even making that comparison. Now, let me tell you about the things that stood out to me while testing. First, the long horizon capability. This is the best part of the model. It has great ways to use the tools that it already has available. For example, it can use Chrome CLI to check its results automatically in the browser. Look at what's rendering, and then go back and fix stuff on its own. I didn't set anything special up for that. It just figured out that the tool was there and used it. Most models don't do this unless you explicitly tell them to.

Second, it doesn't give up. Like I said before, it grabs a task by its head and keeps working until it's actually done. On the Gemma, Finituning task, it hit multiple errors, and just worked through all of them without asking me anything. Third, it doesn't overthink. This is a really underrated point. A lot of the recent reasoning models will burn through thousands of tokens thinking about a simple task before doing anything. K3 just thinks the right amount and gets to work. It makes the model feel fast and also keeps the costs down, which is always a big deal for me. One more thing to note is that the model works way better with Kimi CLI. It's very much optimized for that. I tried it in a couple of other harnesses and it works fine but the difference with Kimi CLI is noticeable. So if you want to get the best out of this model when it launches, I'd recommend using it through Kimi CLI. So what's the verdict? If you ask me, Kimi K3 is the real deal. It's third on my leaderboard behind Fable 5 and OPEC 4.8.

But in real-agentic work, it feels better than OPEC 4.8 and GPT 5.6 sole. The long horizon performance, the tool usage, and the fact that it doesn't overthink make it one of the most practical models I have tested. If Moonshot prices this anywhere near their previous models, this is going to be a really good option for sure. Again, I don't know yet if it's going to be open or not, I'll pin a comment with that info as soon as I have it, overall it's pretty cool. Anyway, let me know your thoughts in the comments. If you liked this video, consider donating through the Super Thanks option, or becoming a member by clicking the Join button. Also, give this this video with thumbs up and subscribe to my channel. I'll see you in the next one. Until then, bye!