What you're seeing is from Kimi K3, a

model from Moonshot AI that's open

source and open weight, which means you

could download it onto your own PC. That

is if you can handle 2.8 trillion

parameters. It has a 1 million token

context window, which is massive, and I

want to explore exactly what makes it

special. Let's find out. When I jumped

on Twitter, I saw their official blog

post here as well as other things like

this benchmark showcasing how Kimi K3 is

essentially beating Fable 5 as well as

ChatGPT 5.6 on the front-end code arena.

Heading further down, I can see that

some people like Theo are already pretty

excited to see how it goes in terms of

3D modeling as well as orchestration,

which I'm excited to test out, too.

People are also discovering that it's a

lot cheaper compared to things like

Fable and ChatGPT, which have a

ridiculous pricing when it comes to

token use, and they just seem to use a

lot more tokens now. Here's the official

blog post for Kimi K3, and there's a lot

of information here if you want to learn

more about it, but I'll give you some of

the highlights. Like here are all the

benchmarks for things like coding and

general agent use. And usually Kimi K3

is now in the top three position. It's

even coming in first place for some of

these benchmarks beating out ChatGPT

Soul as well as Fable 5. And I'm really

interested to test this out in the real

world with my own examples. If we scroll

down further, we can actually see that

its token use is a lot more efficient.

Jumping onto the pricing, we can see

that 1 million tokens per input is $3,

and 1 million per output is $15. Now, if

we compare this to something like Fable,

there's going to be a huge three to five

x difference in terms of pricing that

you'll actually be able to afford,

especially when you're doing a

long-running tasks like people are now

with these one props that create entire

games or systems or startups. But here

you can see some graphs of that, and one

of the issues I'm often having is when

I'm using Fable 5, I often run out of

that 4-hour limit, and I swap to GPT

5.6, which then I run out of that limit.

So, having a model that actually can

keep on running throughout the whole day

would be really useful to me. If you

want to start using it, it's already

available here inside of Kimi code, Kimi

work, Kimi the website, as well as our

API. And when it comes to open source

and open weights, I'm pretty sure it's

the biggest model at 2.8 trillion

parameters. We have the largest behind

it being DeepSeek at 1.6 trillion.

Actually, I double-checked this on

Google, and as far as Google AI says, it

is the largest, which is absolutely

insane. The massive leapfrog it has over

the previous models to a point where I

really wonder just how many Mac minis is

required to be able to run this thing.

There is some more information here in

terms of how it actually operates under

the hood. What I'm personally interested

in is how it'll go for coding, but as a

part of that, I've heard that one of the

greater features that it's now capable

of doing is game dev and 3D design

because it's meant to have very strong

3D reasoning, coding, and vision

capabilities. And there are some great

examples to showcase this. Let's have a

look at these. The first example from

Kimi is this one here that looks like

something out of a Red Dead Redemption,

an open world with horse riding, rain,

water, trees, and much more. What it's

doing under the hood is using a

combination of vision and a loop, as

well as coding, to continuously check

screenshots and essentially build an

array on a game to get it to a point

where it creates something like this

Cyberpunk Spider-Man-like game. And it's

because these models are larger and

better with vision that it can do this.

Here's another example from someone on

Twitter who replicated this 3D game that

pulls inspiration from Paper Mario. And

it's doing a lot from not only flipping

the character, but even making the trees

transparent when the camera's too close.

Higgsfield AI have made their own

comparison between GPT 5.6 Soul and Kimi

K3 for a video generated with Seed Dance

2. And you decide which you think is

better based on the same prompt

essentially generated by each model.

Bridge Bench also has this a lava lamp

test, which Kimi K3 did probably one of

the best jobs I've seen so far, even

beating the example from Claude Fable 5.

Kimik A3 is also pretty good at creating

infographic style reports and

presentations. Not something I do as

often, but what I really am excited to

see is how it goes at creating

dashboards and UI elements because the

examples shown were really interesting,

such as resizable cards, bento grids, as

well as statistics that have their own

unique look and feel because I've been

finding the results from Fable and 5.6

to all look the same. Another point of

note is that it's pretty good at video

editing, from motion design to animation

to video editing in general. You see,

there aren't really any extensions where

AI is jumping into Adobe Premiere or

DaVinci Resolve and really helping you

with the process. I'd love to see where

this goes because right now in coding,

we have agents that plan, generate, and

even test the code for you. But in video

editing, that's not quite there yet. If

you weren't aware, there's also Kimik

Code, which I want to test out with the

new K3 model. Kimik Code is like Claude

Code or Codex, just its own CLI where

you can directly access the model and

it's got its own little harness that

works quite well. Let me jump in here

and select {slash} model and swap to the

latest K3 model. I've got a prompt where

I want to try out a 3D web design to

create a web page for my Sony FX3, which

I'm often filming with, and I'd love to

create one of those animated sequences

where as you scroll through the page,

you get to see the camera unfold into

the 3D elements it's made up I'm also

going to enable yellow mode, which

allows it to keep running without asking

me constant questions or to approve

different tasks. After running for a few

minutes, it generates this here with

this 3D mock-up of the FX3. This is a

pretty impressive. I know that some

models couldn't even create SVG shapes

in the past and now we've come to a

point where they can create 3D mock-ups

just like this. I can expand through

each one of these sections. I can click

through to any parts of the camera and

view more information. And overall, this

is a really engaging and interactive

page. Heading back to Kimiko, I've got

this massive prompt here to create a

Game Boy Advance emulator as a web page.

After running for a few minutes, it

generates this, which is damn

impressive. It's built on top of Vite

and TypeScript and 3.js using real 3D

models, but not only that, it's using a

real emulator. And the game you're

seeing in there is an actual ROM that's

loaded up and playing. If you're

wondering how this is possible, I'm

doing things like jumping onto Sketchfab

to grab 3D assets like this one here of

a space elevator. And then I grab this

URL and jump back into Kimiko. I can

plug in a prompt something like this

where I link it to the Sketchfab 3D

model. And then I ask it to replicate a

scene of a space elevator where it's

going from Earth all the way to space.

Kimiko is using those assets as well as

its better understanding of visuals to

create this design for a website using

things like 3.js and the models

themselves and creating the space

elevator launching into space. Space

itself as it's rising. This really gives

you a taste of what's already possible

as well as what the future might look

like with how quickly these models are

progressing. A big thanks to Kimiko for

sponsoring today's video. If you want to

learn more about them, definitely check

them out in the comments below.

Otherwise, I'll see you guys in the next

one.