First Look at Kimi K3: The Biggest, Smartest Open Weights Model Ever?

[00:00:00]So, Moonshot AI just released Kimi K3, which they're calling open frontier intelligence. It's going to be an open weights model, and it's their latest flagship model in the Kimi line, which is very popular.

[00:00:13]You'll remember I did a first look for Kimi K2.6, and that one performed very well for an open weights model. So, I'm excited to see what the Kimi K3 can do.

[00:00:24]So, we're going to start out just by doing a opening task, then I'm going to come back and try to give you a breakdown of what we learned about it from the tech aspects, and also give you the benchmarks, and then give you a few more tasks at the end. But, I figure we want to start actually messing around with it, not with the slides.

[00:00:42]Um so, you can see Hermes updating in the background, but uh this is going to be the prompt we give it, and it's one that you've seen several times. Uh the prompt is to build a single HTML file for a One Piece meets Star Wars game that has a full-screen hero section with a dark background, rotating 3D sphere in 3.js with a custom GLSL fragment shader, makes it look like a glowing planet. The sphere slowly rotates on its Y axis. On scroll, GSAP animates the sphere scaling down and fading out as text headline fades in. Headline has a framer-motion stagger entrance. Works in Chrome, no NPM, CDN imports only.

[00:01:19]You remember that this was the Kimi uh K2.6 attempt at this task.

[00:01:24]Looked pretty good.

[00:01:27]And at the time, this was actually the best um product from this. Since then, Fable has been produced a better one, as well as uh most recently ChatGPT 5.6 Soul has done a good job with this. But, at the time, this was really good when this was first released. It actually properly did it, where the other frontier models couldn't do it, like Opus at the time, and the GPT model at the time couldn't do it. Uh but, let's see what Kimi K3 does with this.

[00:01:54]Couple months ago, this was a really difficult prompt, but a lot of the more recent models have been doing it really well. So, we may need to make a change if all the models are able to handle this.

[00:02:04]Um so, here we are choosing our model.

[00:02:07]Go down to Moonshot Kimiko 3, you can see in new portal. And you can see the prices here. Right now, it's $3 input, $15 out.

[00:02:17]Uh which puts it roughly around here.

[00:02:19]GPT-5.6 Terra. So, like the mid uh mid-stage level for OpenAI. So, it's not cheap. It is expensive, but it is also open weights. It is available right now.

[00:02:32]Uh 23 here.

[00:02:35]Okay.

[00:02:36]Let's go into Hermes chat.

[00:02:38]And I'm just going to paste in this prompt.

[00:02:41]Okay. You can see the Kimiko 3 here.

[00:02:44]Okay. So, this is what we get.

[00:02:47]You see Grandline Prime, they're calling it.

[00:02:49]One Piece Star Wars Outer Rim.

[00:02:52]So, the sphere itself looks great. Um this is one of the better ones, better versions of this.

[00:02:59]Looks has more realism to this, kind of like a looks kind of like Earth, like there's almost clouds beneath it.

[00:03:06]Very, very good job on the sphere itself.

[00:03:10]So, let's scroll down. This is the real test.

[00:03:13]It's getting smaller, kind of fading out, and nice. You got the that stagger entrance of the headline.

[00:03:21]Set sail for the stars, become the pirate king of the galaxy.

[00:03:25]Uh so, it did a really good job. I would say visually it's much better than uh what it did in the previous version.

[00:03:33]Uh a lot more, I don't know, polished, I would say.

[00:03:37]And it achieved everything that I asked it to do in terms of technical requirements. So, pretty pretty good job.

[00:03:43]And I may need to switch up my kind of go-to prompt because the last couple models have uh did a done a really good job on it, so I'll do some searching for some more difficult kind of front-end coding tasks.

[00:03:58]Okay, so that's our first little task done. So, now let's go back and kind of talk about the model a little bit.

[00:04:03]It is a 2.8 trillion parameter model. It's the first open three trillion class model that has been released. It has 1 million context window, which you saw before.

[00:04:16]And it has a new architecture that's really designed for long-running coding research and knowledge work.

[00:04:22]But the honest framing is important.

[00:04:23]Moonshot is not claiming K3 beats every closed model.

[00:04:27]In its own announcement, the company says that K3 is still trails Claude 5 and GPT 5.6 Soul overall.

[00:04:34]So, the interesting question here isn't whether or not this is the biggest open model. It obviously is.

[00:04:40]You cannot run this locally. Even if it is open weights, it's not feasible.

[00:04:45]Even if you have a very large, really good setup with hardware and GPUs.

[00:04:50]The question here though is whether all that size actually turns into better finished work. So, let's get into it.

[00:04:57]So, Kimi K3 is not simply Kimi K2 with more experts bolted on.

[00:05:03]The first new piece is Kimi Delta Association, KDA here.

[00:05:07]This is Kimi's efficient attention architecture for handling very long sequences without paying the full cost of standard attention across every token.

[00:05:16]Then there's attention residuals.

[00:05:18]Instead of every layer just accumulating everything from the previous layer, the model can selectively retrieve useful representations from the different depths.

[00:05:27]Underneath this is stable lean MoE.

[00:05:30]So, K3 has 896 experts, but only 16 are active for a given token. So, that's why the total model itself is enormous, but the active path stays very sparse.

[00:05:42]So, Moonshot says that the combination of architecture, routing, training, and data produces around 2 and a half times better overall scaling efficiency than K2.

[00:05:52]So, that efficiency claim matters more than the raw parameter count. A giant model is only useful if you can train and serve it without the cost just exploding.

[00:06:00]So, the central pitch of K3 is that long horizon work can be done with minimal human oversight.

[00:06:07]Moonshot says it can navigate very large repositories, coordinate terminal tools, sustain engineering sessions, and use screenshots as part of the coding loop.

[00:06:15]So, it can build something, look at the rendered result, and then revise it instead of stopping after it writes the source.

[00:06:22]And they tested GPU kernel optimization with up to 24 hours per run.

[00:06:28]During K3's own development, early version reportedly handled most of the team's kernel optimization work. You can see this kind of broader pattern across the entire release. It's more focused on kind of taking ownership of an entire difficult project and just keep running until you complete it.

[00:06:44]They talk about a bunch of these different kind of wild case studies.

[00:06:48]And I won't go in detail with all of these. They're on the tech blog if you want to read them. They're pretty interesting though. Um building a mini Triton GPU compiler, a chip built by the model, a procedural 3D open world using 3.js, uh web GPU, a bunch of other assets, and then computational astrophysics.

[00:07:06]It was able to do all of that research and everything that it needed to take, you know, a couple weeks worth of work and put do it in two hours.

[00:07:15]So, let's get into the benchmarks cuz people want to see that, of course. Um there's a bunch in their blog, and you can see in the coding benchmarks here, it's towards the top, right? For most of them. Couple of them losing to GPT-5.6 Soul, couple losing to Fable.

[00:07:29]Um couple of them is actually on the top. It beats Soul and Fable, which is pretty incredible.

[00:07:35]And then for agents, similar kind of setup, right? It lost to Fable and Soul on a couple of these, but it also came up higher than them in like automation bench, browse comp, spreadsheet bench, too.

[00:07:49]And visual agents as well.

[00:07:51]It was above soul slightly below fable though.

[00:07:55]So it's not the best model out there, but it is close to you know the top models we have access to.

[00:08:01]And considering fable 5 was considered a national threat that had to be kept away from the public. That's pretty incredible.

[00:08:10]And you can look here the cost efficiency is a big factor as well.

[00:08:14]The cost per task here K3 this little red star. You can see it's much more efficient in terms of cost.

[00:08:21]So this is kind of my honest read here where K3 looks pretty strong in terms of the benchmarks.

[00:08:27]Those kind of big wins in coding benchmarks and agent benchmarks, strong spreadsheet and document understanding results, visual coding and long context are core capabilities that are really useful.

[00:08:38]Perhaps is not as strong on kind of really difficult software engineering tasks.

[00:08:44]And GPT 5.6 stays ahead on several reasoning and multimodal tests.

[00:08:49]Which moonshot itself kind of admits there's a UX gap versus these top models.

[00:08:54]But still very impressive.

[00:08:56]So I mentioned this before but price and access only 30 cents for cash hit input, $3 for normal cash hit miss input. And then $15 output. Like I said it's kind of on par with GPT 5.6.

[00:09:10]Terra the mid mid-range one.

[00:09:13]So there's the facts of the the model itself.

[00:09:16]Bottom line is that Kimi K3 is not automatically the best model just because it has the most parameters.

[00:09:21]What makes it really interesting is that has a combination of open 3 trillion class model, a million token context window, and genuinely new attention architecture.

[00:09:30]Native multimodality and a focus on work that can run for hours or days.

[00:09:35]So we did one task and let's do another.

[00:09:38]Okay, so the next task I'm going to give it is going to be one based on what we saw in the slides there.

[00:09:44]Uh, this one, procedural 3D open world.

[00:09:46]I want to see if we can create an open world cuz I've had um I had a task that I did with Opus and then the GLM 5.2 model where they created this 3.js kind of open world using some very basic assets. So, I have this prompt um and this is basically saying, you know, "You are in a fresh directory. The only supplied material is assets con- containing fantasy and game assets." And those are like here.

[00:10:17]There's some You can't read these, but these are all kind of 3.js and some reference photos. This was for like a original One Piece game I was trying to make. And this is probably the most useful for them. Like it has all these textures and stuff um that it can use for buildings or like the ground.

[00:10:34]So, there's a lot of assets there that it can use pretty well.

[00:10:38]So, that's kind of the idea. Create this kind of 3D fantasy world uh mainly in just 3.js.

[00:10:46]And there's a couple different requirements for it. Uh, movement controls, three points of interest, one animated element, let the user click or inspect at least one object.

[00:10:55]And yeah, so this is the prompt I'm going to give it. See what we can build and if it's better than what we got for Opus 4.8 and GLM 5.2.

[00:11:06]Okay, so once again, Hermes here.

[00:11:09]I'm going to paste in that prompt. We're in the right directory. This is just a blank directory except for that assets folder.

[00:11:17]Let's see how it does.

[00:11:18]So, it finished it. It took around 7 minutes altogether.

[00:11:23]Um I had some connectivity issues. I think a lot of people are trying to use it today.

[00:11:27]Uh, we got it We got it done. And let's see what we got.

[00:11:32]Okay, here we go. Amber Amber Fallvale.

[00:11:36]Click to enter here.

[00:11:39]This guy looks good.

[00:11:41]There's our little fantasy town.

[00:11:43]Did a good job with the the textures here.

[00:11:47]Got a well.

[00:11:49]I click on it. The old well.

[00:11:52]Not a very creative name, but Um Let's see.

[00:11:59]The fallen star crystal.

[00:12:03]And who's this?

[00:12:05]The hooded adventurer.

[00:12:08]So, interesting. They did a good job with light.

[00:12:11]Um like that light effect. I know it doesn't seem like much, but that's pretty good.

[00:12:15]It's coming from the sun.

[00:12:18]The way it balances it out, that looks pretty good.

[00:12:21]Uh so, this is nice.

[00:12:24]Little castle over here.

[00:12:27]They did a pretty good job for just one shotting this. Obviously, if you're going to bake a real 3D world, you could refine it a little bit, but it had all the elements I see.

[00:12:37]You can click on stuff and it gives you a name up for whatever it is.

[00:12:42]There you go.

[00:12:45]And if you don't remember, this was the the Opus 4.8 version of this.

[00:12:56]They have more characters, but I don't know.

[00:13:01]I feel like the houses are are not as well done.

[00:13:05]This water was better, though.

[00:13:08]But yeah, this was the Opus 4.8, so I don't know if you consider this better than what we just saw. Uh give me K.

[00:13:15]K3D I wouldn't necessarily say so.

[00:13:21]Right?

[00:13:22]This one feels kind of richer. I don't know the world.

[00:13:26]It uses a lot more of the textures and stuff, I feel like.

[00:13:30]Little pull right there.

[00:13:35]Okay, and lastly, we're here in the desktop app. I wanted to try this cuz I want to try it with sub-agents.

[00:13:41]They say this Kimi K3 model is really good as an orchestrator model.

[00:13:44]So, I want to give it kind of a a goal where it has to use sub-agents and see how well it how well it can synthesize information.

[00:13:51]Uh so, this is the task. I said, "Spawn three sub-agents to do research on the following topics.

[00:13:56]Uh research important research papers in the last 3 months about one agent memory, two agent tool use, three agent skills. And then, synthesize the information from the sub-agents into a coherent report giving me the top five most important papers ranked in terms of legitimacy and novelness and impact on agents."

[00:14:14]So, let's see uh how it can handle that kind of orchestration role here. And I did this in the desktop app cuz we should be able to see see them do this.

[00:14:30]I'll first load the research skill for proper workflow guidance, and then spawn the three sub-agents in parallel.

[00:14:36]So, we'll be able to see them here.

[00:14:40]And this is obviously using delegate task as the as a tool.

[00:14:45]Let's see.

[00:14:46]Got our three workers now. These are the kind of their their goals they've been given by the orchestrator.

[00:14:52]You could they gave it a lot more context. Like Mike was just saying, like agent tool use. You can see it's very specific about papers about tool learning, tool function calling, MCP, tool use benchmarks.

[00:15:05]Give it very strict um instructions. Do not invent papers or ideas.

[00:15:10]That would be certainly very important.

[00:15:16]So, so far, so good. Let's see what we get back from them.

[00:15:22]Okay, it's chugging along.

[00:15:24]Uh all the sub agents just finished their work.

[00:15:27]So, now it's going to review what came back.

[00:15:30]You can see it thinking here.

[00:15:33]About eight papers from each of them.

[00:15:38]Seven paper papers from the skills agent.

[00:15:43]Now it's going to synthesize them and create the like top five most interesting papers.

[00:15:50]Interesting to see the thinking here.

[00:15:54]Ranking it by legitimacy, novelty, and novelty and impact.

[00:16:00]Interesting. Now it's spot checking the the top five cuz it said in the thinking it said it couldn't confirm or wouldn't necessarily trust what the sub agent said.

[00:16:09]And it would be fairly cheap to confirm the key claims in the five papers it chose.

[00:16:14]So, it's certainly doing more work and thinking through this process more.

[00:16:18]And I only have this on medium.

[00:16:20]Medium thinking.

[00:16:24]Okay, so it verified the top five candidates. They were real pages.

[00:16:27]Everything was matching.

[00:16:29]So, here's the synthesized report it wrote for me.

[00:16:36]And number one was this OSWorld computer use agents on long horizon real world tasks.

[00:16:42]Number two here is agent memory uh characterization and system implications from Stanford.

[00:16:50]And then MR agent skill composer.

[00:16:54]And then reward hacking benchmark.

[00:16:59]So, it's aided a pretty good job of synthesizing all the information.

[00:17:03]Um it gave it a score. It gave it the reason why it did. It gave it a nice little summary here of each of them.

[00:17:11]And it gave kind of this nice little field level trends which are consistent across all three branches.

[00:17:16]Evaluation shifted from it to should it.

[00:17:19]Memory is becoming infrastructure. Scale research is answering its skeptics.

[00:17:26]So, it handled it very well. There are no hiccups. It was pretty fast, I would say, in this one.

[00:17:30]Um it took uh less than 10 minutes altogether.

[00:17:34]Uh so, it did a pretty good job as an orchestrator agent. Um the next kind of test would be like more complicated multi-agent workflows and stuff like that. But for this simple research one, thought it did a pretty good job.

[00:17:48]There you have it. That was my first look at Kimi K3, uh very important model, the largest open weights model uh that we've had so far, and the one that's scored the best in benchmarks.

[00:18:00]So, I'll be interested to see once they actually release the weights what we can glean from that.

[00:18:05]Uh but my first impressions are pretty good. Like it does definitely feel like in the top frontier models in terms of performance, but I'll have to use it more to really really figure it out.

[00:18:15]Um but that's going to be it for this video. Please leave a comment. Let me know your experience with Kimi K3.

[00:18:20]And I'll see you in the next video.

[00:18:22]Thank you for watching.