Kimi K3 Just Matched GPT-5.6 Sol & Claude Fable 5 Performance | Beats GLM 5.2

[00:00:00]Hey guys, welcome back to another new exciting video. Finally, Kimi K3 model is here and here you see that I have tested this model and this is the output from Kimi K3.

[00:00:10]Okay, and now let me show you the output from the GLM 5.2. So, this is the output from the GLM 5.2 and now this is the output from the GPT 5.6 data max. Okay. And uh now the output from this Kimi K3, this is the max version. Basically, if you go to this kimi.com official website and there you will find this K3 max. I tried this one and there is another which is the K3 swarm max. Okay, but swarm is for multi-state workflows batch processing, but this K3 max is enough for making this kind of website. Okay, a single page landing landing page kind of website. So, I choose this one. Now, one thing is that on their official website kimi.com, if you want to use this K3 Kimi K3 model, here you see that there are some free quotas. Okay, and after making this website, this 3D globe website, I found that it is giving me this message. Your free quota is used up. Please refresh this at uh uh 8:12.

[00:01:19]Okay.

[00:01:20]Now, the thing is that whether this Kimi K3 model can beat or can give the similar kind of intelligence like the Febel 5 or the GPT 5.6 all. So, here you see that they have actually compared their model with the Opus 4.8 GPT 5.6 all and Febel 5. So, on this SWE marathon here you see this model is the number one model and not only that on this Kimi code bench, that is their internal benchmark, they are beating this all of the um open AI model and also it is beating the Opus 4.8.

[00:01:56]On this frontier SWE, that is the popular benchmark, they are also you see they have beaten all of these open AI model and also this Opus 4.8. And Fable 5 is the number one.

[00:02:06]So, from all of these this terminal bench, program bench, deep SWE, which is the very popular benchmark and we have already seen that uh GPT 5.5 was the number one model uh for a long time on this deep SWE benchmark and Opus 4.8 was the number two model on this deep SWE benchmark, but this Gemini A3 they have beaten these both two model. And also the number one model number one open source model GLM 5.2.

[00:02:35]Okay, so now we can actually get the idea that how much powerful this model is. And also from this output, this is the output this 3D globe dashboard that it has made.

[00:02:47]Actually, I have not got this kind of output from the GPT 5.6 data also, which is beating the uh Fable 5. Okay, means if you uh if you go to this GPT 5.6 benchmarks, if I write this and I think if I go to images there we'll find Yes. Uh this one. Here do you see that GPT 5.6 data is beating Claude Fable 5.

[00:03:15]But if you see this GPT 5.6 data output and if you compare this one with this Gemini K3 output, there is much difference.

[00:03:26]Okay? And Gemini K3 output is actually much polished. Right? And this output actually matching with the Fable output. Here do you see this is the Fable 5 output on the right hand side.

[00:03:39]And if I play this video, so here do you see.

[00:03:48]This world map and this this sea water and this soil all are actually clearly visible that where it is present and currently they have kept this pricing as the $3 and $15 per 1 million input and output token which is very very less than the Fabel 5 and also we are getting the almost same kind of output like the Fabel 5 and I will say that this output is much more polished and much more cleaner than all of the existing frontier open source and close source model. Okay. So Kimi K3 actually did a great job and if you go to their official blog post then you will see that uh they have implemented a new architecture which is the Kimi Delta attention KDA.

[00:04:34]Okay, previously we have heard about the sparse attention which Minimax implemented MSA Minimax sparse attention but Kimi actually used this Delta attention. Okay, and also the attention residual this both technique they have used and both architecture updates are designed to help information flow more smoothly through longer sequences and deeper models. Okay, and also further increase the sparsity of the mixture of experts means this model is a mixture of expert model and this is a 2.5 trillion parameter model. I think it is 2.8 trillion parameter.

[00:05:13]So uh Here you see with the stable latent MoE framework the model effectively activates 16 out of 896 uh 896 expert together with improvements in training methodology and data recipes.

[00:05:31]The structural advances give Kimi K3 roughly 2.5x the overall scaling efficiency of Kimi K2. Now here they have mentioned some of the terms like this stable latent MoE. What is this? So if you see this is the actually this is the actually stable latent MOE, left-hand side upper one.

[00:05:53]This is the stable latent MOE, okay? And below here you see, this is the Kimi Delta attention, KDA.

[00:06:00]Now, what is this stable latent MOE?

[00:06:02]What is the purpose? It activates only a small fraction of its total 2.8 trillion parameters per token while maintaining the stable training, okay? Now, what is the purpose of this Kimi Delta attention, KDA? Um it actually reduce the KV cache size and also making the very long context, like 1 million context it supports.

[00:06:29]It keep it as a very practical, means you can really use the 1 million context window in an efficient way.

[00:06:35]And another term which is the attention residual, okay? What is the meaning?

[00:06:40]Means what is the purpose? So, it actually helps to reuse the attention computation um again again without recomputing the similar relationships, okay? And this is the actually architecture and here you see this is the embedding and after that there are some blocks and after that we have this KDA and stable latent MOE and after that we have this output section, okay?

[00:07:02]Actually, Kimi K3 is more focused on software engineering on long context document understanding, long context code bases, tool usage, multi-step reasoning and also the autonomous coding agent. Here you see on coding, it has a strong long horizon coding capabilities with minimal human supervision. It can sustain long-running engineering tasks, understand and work with large code bases and coordinate terminal tools. That's why here you see, they are actually beating the GPT 5.5 and Opus 4.8 on this Deep SWE, okay?

[00:07:37]And also this knowledge specific work, it advances the end-to-end knowledge work beyond public benchmarks Gemini K3 Max version also shows consistent gains in our internal evaluations. These evaluations reflect a recurring task pattern and challenges from real user agent collaboration workflows. And whether this model is an open model, see, I have not found any hugging face page for this, but they are saying that Gemini K3 open frontier intelligence maybe in the coming days they will publish this model as the open source.

[00:08:10]source. Yes, here you see the full model weights will be released by July 27, 2026. Further details on the architecture, training, evaluation will be released alongside the Gemini K3 technical report. So, that is why I was saying that currently they have just released the model on their official website and in their API, but um Here you see, we are currently working closely with the inference partners and open source maintainer to align the technical details and ensure a reliable rollout across the ecosystem. So, they will make this model as the open source.

[00:08:45]But for that we have to wait till the July 27.

[00:08:48]Okay.

[00:08:49]So, yes, this is the thing guys. I think I have discussed all of these important points.

[00:08:56]And these are the some examples that they have given like for the game dev and digital creation.

[00:09:02]Here you see how it is performing.

[00:09:07]And see here this this this one.

[00:09:10]This is actually actually giving the top frontier closed source model level of performance, right?

[00:09:19]And let me check this one.

[00:09:22]Let me check this.

[00:09:25]This is actually quite huge achievement. They have done a huge achievement, guys.

[00:09:29]On the code arena of this web dev where they evaluate the capability of the models on front-end development task and um also the agentic coding workflows that requires the multi-step reasoning and tool use. On that benchmark, here is Sekimi K3 is the number one model, and it has uh beaten this uh GPT-5.6 all XI and also Claude 75, which is really, really a great achievement.

[00:09:55]So, yes, guys, these are the things uh that is important to know. I have shared all of that, and uh when they will make this model as open-source and when when more inference provider will uh actually make it available on their platform, I will make more videos on this. And uh but overview, I think you have got, and um this model actually giving the great capability like the top models, top closed-source model.

[00:10:21]So, yes, if you found this video helpful, don't forget to subscribe this channel, don't forget to like this video also. See you guys in the next video.

[00:10:26]Thanks for watching. Bye-bye. Take care.