Modern Creator
Why Modern Creator?
Pat Simmons · YouTube

Kimi K3 Is Here! (Better Than Opus 4.8?)

Ten identical builds, five models, blind-ranked before the reveal — a real-world stress test of Moonshot AI's new open-source model against GPT-5.6 Sol, Opus 4.8, GLM 5.2, and its own predecessor.

Posted
yesterday
Duration
Format
Review
hype
Views
13.9K
571 likes
Big Idea

The argument in one line.

Kimi K3 posts strong benchmark numbers against frontier models, but a blind ten-build comparison shows it's inconsistent at agentic coding while punching above its weight on knowledge-work tasks like writing and motion graphics.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You're deciding whether to add an open-source model like Kimi K3 to a coding stack that currently runs on Claude, GPT, or Gemini.
  • You're not a developer but want to know whether cheaper open-source models can already handle everyday writing and knowledge-work tasks as well as frontier models.
  • You're comparing AI subscription tiers and want a real account of how fast a cheap plan's usage limit gets hit under actual agentic coding load.
SKIP IF…
  • You want a controlled, statistically rigorous benchmark study — this is one person's subjective blind ranking across a small sample of builds.
  • You only care about published benchmark scores, not how a model performs on real generated apps and games.
TL;DR

The full version, fast.

Moonshot AI's Kimi K3 claims frontier-level coding performance at a fraction of the cost of GPT-5.6 Sol, Opus 4.8, and Fable 5. This video tests that claim with ten identical builds — 3D gadgets, a weather app, a fighting game, a Counter-Strike clone, motion graphics, biography writing, and landing-page copy — run in parallel across all five models and blind-ranked before reveal. Kimi K3 showed a dramatic jump over its own predecessor K2.7, but placed inconsistently against frontier models on coding tasks (first once, fourth or fifth several times) while outperforming them on knowledge-work tasks like biography writing and motion graphics. The cheap Moderato subscription tier also burned through its usage limit after just two builds, forcing a mid-video upgrade.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 06:41

01 · Intro & methodology

States the premise (Kimi K3 vs. open-source and frontier rivals), explains the agent fan-out method across six terminal windows via Claude Code, and covers access routes (OpenRouter, subscriptions, kimi.com CLI login).

06:41 – 09:05

02 · Interactive 3D Gadget

Five models build a rotatable 3D 'command center' device. Kimi K3's build wins the round with working power sequence, sound effects, and dynamic lighting.

09:05 – 11:39

03 · Weather App

Models build a weather forecasting app with real data and a custom short-term forecast model. Several converge on near-identical blue gradient designs; GPT-5.6 Sol ranks first, Kimi K3 ranks near the bottom on aesthetics despite working functionality.

11:39 – 14:10

04 · MP3-to-MIDI Converter

Viewer-suggested build: an app that detects notes/pitch/timing from an audio file and converts it to MIDI. One model fails to load the file entirely; Kimi K3 and K2.7 produce near-identical, hard-to-distinguish results.

14:10 – 16:37

05 · 2D Fighting Game

A playable Street Fighter-style browser game with combos and health bars. Kimi K3 shows the biggest generational jump of the video over Kimi K2.7, described as 'night and day.'

16:37 – 18:43

06 · Luxury Watch Landing Page

A photorealistic, rotatable 3D watch hero for a product landing page. Fable 5 and GPT-5.6 Sol impress most; Kimi K3 underperforms specifically on this 3D task despite strong results elsewhere.

18:43 – 21:34

07 · Counter-Strike 2 Demo

A playable browser clone of Counter-Strike 2. Fable 5 and GPT-5.6 Sol rank ahead on physics and controls; Kimi K3's build has strong graphics but confusing, inverted movement controls and ranks fourth.

21:34 – 25:31

08 · Isaacson-Style Biography

Models research and write an opening biography paragraph about the host in Walter Isaacson's style, with no input but his name. GLM 5.2 and Kimi K3 rank first and second, beating both frontier models tested.

25:31 – 27:40

09 · Motion Graphics Explainer

Using a motion-graphics skill, models generate an animated explainer video about Kimi K3 itself. Kimi K3 produces the most kinetic, professional-feeling result; GLM 5.2 fails to generate output at all.

27:40 – 32:45

10 · Workshop Landing Page Copy

Models rewrite copy and design for the host's real AI Boot Camp landing page. GPT-5.6 Sol and Kimi K2.7 rank best; Kimi K3's copy is judged the weakest of the entire video, ranking last.

32:45 – 34:48

11 · Tarantino Trailer

Models write and fully shot-list a 30-second Quentin Tarantino-style trailer, later generated as video with Sora/Seedance. GLM 5.2 ranks first; results are called the least conclusive test of the video.

34:48 – 36:06

12 · Final Tally & Outro

Recaps Kimi K3's placements across all ten tests, concludes it's a strong pick for everyday knowledge work at a much lower price than frontier models even though it was inconsistent on hard coding tasks.

Atomic Insights

Lines worth screenshotting.

  • Kimi K3 beat GPT-5.6 Sol and Opus 4.8 on several published coding benchmarks, but real-world build tests told a much more mixed story.
  • Across ten head-to-head builds, Kimi K3 finished first once and placed fourth or fifth in roughly half of the coding tests.
  • On knowledge-work tasks like biography writing and motion graphics, the open-source models beat every frontier model tested.
  • A $19/month Kimi Moderato subscription hit its usage limit after only two agentic coding builds, forcing an upgrade to the $39/month tier mid-session.
  • In a blind biography-writing test, the two open-source models (GLM 5.2 and Kimi K3) ranked first and second, beating GPT-5.6 Sol and Opus 4.8, which finished last.
  • Kimi K3's jump in build quality over its predecessor Kimi K2.7 was large enough to be called 'night and day,' especially on the fighting-game and 3D-gadget tests.
  • Kimi K3's landing-page copywriting was the weakest showing of the whole video, ranking fifth out of five models on that task.
  • The testing method: fan identical prompts out to Claude Code across six terminal windows wired to different backing models, then blind-rank the outputs before revealing which model made which.
  • Running the ten-build comparison cost roughly $65 in a single day, split between subscription fees and pay-per-generation API costs.
  • Kimi K3 was not yet available on OpenRouter at review time, requiring either kimi.com's web UI or a separate CLI login flow to test it.
  • A model can render a visually striking interface with broken core functionality and still out-rank a plainer, fully-working build in a blind test.
  • Multiple models independently converged on nearly identical blue-gradient weather-app designs, showing how uniform AI-generated visual taste can get across providers.
Takeaway

A blind ten-build test separates model hype from reality.

BAKEOFF LESSONS

Blind, task-by-task comparison across ten real builds shows Kimi K3 beating every frontier model at writing and motion graphics while trailing badly on coding tasks that demand precise controls.

02Interactive 3D Gadget
  • Kimi K3 produced the most polished result in the 3D-interactivity test, including a working power sequence and sound effects a step above its rivals.
  • The same model's tactile 3D generation quality doesn't guarantee flawless controls — a dial that spun the wrong way or backwards text still made it into a top-ranked build.
03Weather App
  • Multiple models defaulted to nearly identical blue-gradient interfaces, showing how easily AI-generated design converges on the same generic look across providers.
  • Kimi K3's own weather build worked correctly but lost the ranking purely on aesthetic choices, a reminder that 'good AI output' is judged on function and taste together.
04MP3-to-MIDI Converter
  • A model that fails to load the test input outright loses to a mediocre model that at least completes the basic task.
  • Kimi K3 and Kimi K2.7 produced nearly identical results on this task, making it hard to tell whether a version upgrade actually improved this particular skill.
052D Fighting Game
  • The jump in build quality between Kimi K2.7 and Kimi K3 was large enough to call 'night and day' — worth re-testing an open-source model after a major version bump rather than judging it by an old version.
  • Confusing control schemes and unreadable UI text sank otherwise-playable builds, showing usability details matter as much as raw feature completeness.
06Luxury Watch Landing Page
  • 3D rendering and dynamic lighting quality has become table stakes across both frontier and open-source models — a differentiator from months earlier is no longer a differentiator.
  • Kimi K3 underperformed specifically on the 3D product-hero task despite doing well elsewhere, showing model strength doesn't transfer evenly across similar-looking visual tasks.
07Counter-Strike 2 Demo
  • A visually impressive build can still be unusable if basic controls are inverted or unresponsive — graphics quality and playability are separate axes worth judging separately.
  • Across three tests of the same prompt, the paid frontier models consistently outranked the open-source models on physics and control feel.
08Isaacson-Style Biography
  • Given just a name, a capable model will research the person online and cite sources before writing — a useful check on a model's research quality, not just its prose.
  • In this writing test the two open-source models beat every frontier model tested, flipping the usual assumption that frontier automatically means better writing.
09Motion Graphics Explainer
  • Kimi K3 produced the most kinetic, professional-feeling motion graphics of the group, suggesting design/animation tasks are a genuine strength worth testing for your own use case.
  • One model failed to generate output at all on this task — a reminder to budget for outright failures, not just quality differences, when piloting a new model.
10Workshop Landing Page Copy
  • Kimi K3 produced the weakest copywriting output of the entire video, a reminder that a model excelling at code or design doesn't guarantee it's strong at persuasive writing.
  • AI-generated marketing copy has developed its own detectable cliches (em dashes, the word 'quiet,' overlong subheadlines) worth watching for if you're using AI drafts for real marketing copy.
11Tarantino Trailer
  • Even a model that struggles with rigid coding tasks can produce strong creative writing and shot-listing output, so don't judge a model's creative potential by its coding benchmark score.
  • Garbled or random output showed up even in top-ranked creative generations — always review raw AI creative output closely before treating it as final.
Glossary

Terms worth knowing.

OpenRouter
A single API that routes requests to many different AI models from different providers, letting a developer swap models without rewriting code.
Agent fan-out
Running the same prompt simultaneously across multiple terminal sessions, each wired to a different AI model, so the outputs can be compared side by side.
Blind compare
Evaluating multiple AI outputs without knowing which model produced which result, then revealing the source afterward to avoid bias in the ranking.
SWE-bench / Terminal-Bench
Standardized test suites that score how well an AI model completes real software-engineering and command-line tasks.
GDPval
A benchmark that scores AI models on realistic knowledge-work tasks rather than pure coding ability.
Resources

Things they pointed at.

00:30toolAgent fan-out skill (GitHub) ↗
01:00toolOpenRouter
09:40toolKimi CLI (kimi.com/code)
25:31toolHyperframes motion-graphics skill
27:40productAI Boot Camp waitlist ↗
36:06linksimmonsbench.com/kimi-k3 (full generations + prompts) ↗
Quotables

Lines you could clip.

09:05
“Kimi K3, you have officially got my attention.”
direct-address verdict line, punchy and standalone→ TikTok hook↗ Tweet quote
11:00
“I just every time I see blue now, I just I'm triggered.”
funny, relatable reaction to convergent AI design choices→ IG reel cold open↗ Tweet quote
30:48
“Stop renting your operations. Start owning them.”
AI-generated marketing line strong enough to stand alone as a pull quote→ newsletter pull-quote↗ Tweet quote
33:20
“That was really quite bad though, if we're being honest.”
blunt, unhedged verdict on the model's weakest showing→ TikTok hook↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogy
00:00Alright. Well, model releases keep on coming. This time out of the open source world, Moonshot AI has just dropped Kimi k three.
00:07And on the benchmarks, it's beating its open source competitors and even going toe to toe with some of the best frontier models out there. But you know this isn't a benchmarks channel. So in this video, we're putting Kimi k three through the ringer with nine builds.
00:18Everything from games and three d apps to slide decks and creative writing. We've even got a Quentin Tarantino directed movie trailer. Yeah.
00:25We're getting a little crazy with it. So by the end of this video, you will see exactly where Kimmy k three lines up against the Frontier models and answer the question, is this the new open source king? So let's find out.
00:34Getting right into the builds, we're gonna kick off or fan out of a bunch of different building agents to start running through these tests. So I'll just c d into my coding folder here, and then I'm gonna do all of this in Claude code just because it's the easiest way to run all of these across all these different models.
00:51And then I just have this command here that I'm gonna paste in, and this is just telling Claude to, across a bunch of different models, run these builds. And I really wanted to put this to the test across Frontier and these open source models, so we're gonna be testing Kimi k three against g l m five two, against Kimi k two seven, its predecessor, and then Opus four eight because I'm really curious how it holds up against Opus, and then GPT five six Soul as well.
01:15And then we may do a couple Fable tests. I'm really running out of usage credits quick on Fable. I don't know.
01:20Wanna pay extra for those. So maybe for a couple, we'll see how it compares to Fable as well. I wish I had more usage credits, but we'll at least get a sense of how it compares to the other Frontier models as well.
01:30And like I've done previously, I will include this fan out skill. In a second, you'll see six I term windows pop up running all of these different builds. But the way I'm accessing these models is through, first, OpenRouter where available.
01:42So probably familiar with OpenRouter. If you've watched any of my videos, you've seen me use this in the past. So we're using the OpenRouter API key for GLM 5.2 and Kimi k two seven.
01:50I'm using my subscription for g p t five six and Opus. And then on the k three side, this is not available yet in open router. What you need to do is either go to kimi.com and you can interact with k three here.
02:03You can see it's available on the model picker. You'll definitely need to sign up for a subscription. You'll probably get, like, two generations before you'll run out of the free credits.
02:11To sign up for a subscription, you just go here, you go upgrade plan. Uh, you can see I'm on the moderato right now monthly. I'm assuming I'm gonna hit usage limits already though, so I might upgrade for this demo.
02:23But that's how you do it in kimi.com. And then what you can do as well, this is what I'm doing is I'm using the CLI because I prefer to have access to all of these in terminal. So what I'm doing is Kimi dot com slash code, and you can actually go to console.
02:37And once you sign up for subscription plan, you could just use your subscription right here. Then all you need to do is just run a bash command. It's just exclamation Kimi login, and then it will just run the authentication, verify, and you're in.
02:48And these iTerms popped up on my screen, so we are in progress on all of these different builds. We're gonna start with coding, then we're gonna move into knowledge work. So I'm doing six AI builds here.
02:58One is a three d gadget. I'm gonna explain all of these, and I'll show you the prompt here in a second. Once these builds complete but that's the first one.
03:04The second one is a weather simulator. The third is an m p three to MIDI visualizer. The fourth is a fighting game.
03:10The fifth is a Counter Strike two demo. Yep. We're running that one back.
03:13And then the sixth is a landing page for a luxury watch product. Okay. So all of these have completed, and we're gonna go to Simmons Bench here in a second and look at these.
03:21However, what I noticed is already after only two builds just refresh this. Himi k three was only able to get two builds done with my Moderado $19 a month subscription plan, and I'm already hit my usage and it resets in two hours.
03:36And I'm at 20% usage. So that is interesting. I mean, it's they have a lot of different subscription plans, they probably just try to get you in on this cheap one.
03:45I don't know if I'm gonna wait two hours or just upgrade my plan. But, Kimmy, as a daily driver, I'm already realizing that you'd probably need, I guess, something like the Allegro to be regularly using this. Gosh darn it.
03:57Okay. So I just upgraded to the $39 a month plan. Let's see how much usage that will get me.
04:03But, uh, this would be a good time to say please like this video, subscribe to the channel if you haven't already. It helps me out a ton and may help in the slightest of ways subsidize some of these AI subscription costs that I'm paying hundreds of dollars a month for.
04:15So thank you for doing that. And then let's get back to the builds. And while we wait on those k three builds to finish up, let's just quickly talk about the benchmarks.
04:24I promise this would be really quick just so we can draw some sort of comparisons. You know these open source labs are notorious for benchmarking, so take this with a grain of salt. Hopefully, the tests that we run will be more indicative of this than the actual benchmarks, but here's what we're looking at.
04:38First, coding, DeepSUI, Kimi K coming in third behind Fable five and GPT five six Sol, way ahead of Opus four eight, way ahead of GLM 5.2, which is interesting. Frontier SWE, they have it coming in second only behind Fable five and a huge jump above five six SOL.
04:55Then on Kimi code bench two point o, which wasn't familiar with that one, they are certainly not bench maxing their own benchmarks. It comes in second. Well, you know, you gotta hand it to them.
05:03They they could've just ranked themselves in first. Behind Fable five, Terminal Bench two one, ahead of Opus and Fable, and only behind Soul. Program Bench, not familiar with that one, but it's at the top.
05:15And then Suite Marathon, I'm not familiar with Suite Marathon either, but it's also at the top. I mean, just from the coding benchmarks, this is not only just crushing Opus four eight, but also giving Fable and Five six Soul a run for their money with an open source model. We will see though.
05:28I mean, like I said, benchmarks only tell us so much. And then we have general agents, uh, not familiar with most of these, but let's just talk about these quickly. I do know GDP val though.
05:38This is more for knowledge work. You can see Kimi k three coming in third here. There's quite a bit of a differential between Fable five and five six soul and Kimmy k there.
05:46Uh, and then we have a bunch of these other ones, not really familiar with these. I mean, you can see it coming in at the top with a lot of these. Then visuals too, not familiar with this.
05:55I don't even know why I'm reading these out loud. I should not be trusted with naming off any of these benchmarks, so take these with a grain of salt. I I don't know what I'm reading.
06:01I literally just read these out loud so people can see how it compares. All I care about though is getting back to the build, so that's enough benchmark talk. Let's check-in on k three and see how it's doing.
06:10Alright. I just I hate the rate limits once again. We got about four more builds done, so it's still not quite done, but I am just gonna preview the builds that have completed with k three and the other models.
06:21We'll do a comparison. I have a couple others running. What I did, I just switched to the API key.
06:26So I'm just running through their CLI with the API key now, and I'm just gonna pay for each generation because I cannot upgrade again to the $100 month plan. But with that, let's go to Simmons Bench and at least compare some of these builds just to start. So we've got a couple of these ready to go.
06:41First up, build an interactive three d command center gadget. Reminder that all of these prompts will be available in a separate blog post, and you'll be able to interact with and see all of these generations live. But let me just quickly gloss over this, build me a visually stunning, fully interactive three d hardware gadget that runs entirely in the browser, a rugged desktop command center device I can look at from any angle and actually operate.
07:05The idea here is in their demos, in that little trailer that they had, they're doing a lot of good three d generations. So I wanted to put it to the test, like, is where a lot a lot of these you'll see a couple more of these where three d is a big part of it because they emphasize this so much in their trailers. I really wanted to see how it stacks up against the other models.
07:23And just like we've done in the past, I'm gonna look at each one of these generations and then reveal which model is which. So let's take a look at this first one.
07:30Click in here. Okay. Click this.
07:32Oh, look at that. Alright. I mean, not bad.
07:34Oh, alright. Yeah. Kinda hard to text a little backwards, but we got some nice dynamic lighting here.
07:39This this is a little off. This little dial here. Yep.
07:42It's all over the place. Kinda turns around. Oh, but I like the the glare.
07:46Okay. Not bad. Not bad.
07:47Take a look at this one. Okay. A little glitchy.
07:50You can at least read the text. Okay. You can change this even though it's completely wrong.
07:55Click these. Okay. I can click these, but it's just all over the place.
07:58Okay. Oh, power's on, power's off. Oh, look at that.
08:00Nice little power on, you can't really see it. Alright. That one's I'd say that one's a little bit better even though it's not great.
08:06Third one. Alright. We can read it.
08:08Nice. A good this is a good sign. Okay.
08:10Okay. Oh, okay. Alright.
08:11I got it. Sorry. I just need to go up.
08:13Alright. That's really good. That's really realistic.
08:15I got power. Oh, look at that. Oh, nice.
08:17We have a nice little boot up. Okay. This one's good.
08:19Look at that lighting. Oh, look at this. It's even got sound effects, backlight, net.
08:24I don't know what that does. That's a little weird. Sends.
08:26I mean, this one's the best one far. And then we got this last one. Okay.
08:31Alright. Yeah. Readable.
08:32Very similar. Okay. Nice.
08:33Dial works kind of kind of. Okay. Kinda hard to move this.
08:37No sound effects unfortunately, but it does something. We got a nice little backlight. We got a nice little mode.
08:42Power. Nice little power on. Alright.
08:44That one's definitely number two. Alright. Let's see.
08:46Is this Kimmy? No. Alright.
08:49Kimmy k three off to a good start. Uh, Opus4 8? No.
08:53GPT56Soul. Alright. Wow.
08:56Alright. Alright. We're we're we're cooking.
08:58Okay. Opus48 and then KimmyK27. Alright.
09:01Alright. KimmyK3, you have officially got my attention.
09:05Next up, we have build a weather app with its own forecast model. Quickly gloss over the prompt, but it's build me officially stunning weather forecasting web app for North America, pulls real data, and produces its own short term forecast, not just a wrapper.
09:20Requirements, not gonna read all of these. Again, all these prompts will be available to you in the post below, but let's take a look.
09:27Alright. Number one. Uh, okay.
09:29We just have a search. So Toronto. Okay.
09:33Pulling live weather. Oh, okay. Yeah.
09:34Nice little nice little interface. I don't love the emojis, but and this is kind of broken, but some good colors here, some good design. We got the little sun in the background.
09:43You know, it's not bad. It's not bad. Okay.
09:45That's number one. Number two, this already looks way too AI looking. Okay.
09:50Where in North America? Okay. Denver.
09:52Alright. Yeah. Yeah.
09:53It's not bad either. I don't like the blue gradient, but the chart looks better. These charts are actually not broken.
09:58And then we get this nice little thing here. Okay. Pretty good.
10:01I would say one's better than this one. Okay. And then let's look at this.
10:04Alright. I just every time I see blue now, I just I'm triggered. But, I mean, this interface looks pretty good.
10:10Chart kinda broken. That looks pretty good. And then we can click Chicago show up.
10:15Alright. Yeah. Pretty good.
10:17Okay. I would say that one's best yet. And then we have this classic cream.
10:21Somebody in the comments on my five six soul video said background cream color is the new purple gradient and I could not agree more. Now I I can't unsee that now, so let's do I need to give location permissions. That's fine.
10:33Uh, let's do Mexico City. Okay.
10:36Weird icons. This is actually a good test, uh, because none of these really did all that well. We okay.
10:41We have some nice little moving SVG. This one's definitely the best, uh, but you can see it's still kind of broken. It's very readable chart.
10:48Okay. It would have been nice to have, like, a little hover state on this chart here, but that's fine. Same with this one, but at least it's, like, not broken.
10:55Um, Yeah. Okay. So let's see.
10:56I wanna see, like, another one. Uh, New York. Yeah.
10:59Okay. Yeah. It's nice, like, move.
11:01There's actually some animation with each city. Okay. Yeah.
11:05I mean, this one is definitely the best. K three again? No way.
11:08Five six soul. Okay. This is my second favorite.
11:12Opus four eight. Okay. Actually, this one is pretty good.
11:14Is this k 3? Ah, K 27. Alright.
11:16K 3. Damn. Alright.
11:18Alright. Fair enough. Uh, so here's here's k three.
11:21You know? Yeah. Like, this wasn't bad.
11:23The chart was actually working, but I just do not like these frigging blue gradients. So I'd say 56 soul one opus two. I might even say I prefer K 27 though just because it even though, like, had this nice little background.
11:37So there is building a weather app. Next up, we've got an m p three to MIDI converter. And I'll be honest, I don't even know what this one is.
11:43This was an idea that someone left in the comments from my five six soul video, I believe, which thank you everyone for dropping in those comments. Please, if you have any ideas for future model release demos like this and tests, please let me know in the comments. I really appreciate it, and it gives me a ton more ideas, and we can really, like, stress test this because I'm constantly running out of ideas as these models get better.
12:00Anyway, so an m p three to MIDI converter. Here's the prompt. Build me a browser app that converts an audio file into a MIDI file by actually detecting the notes, pitch, timing, and duration, and lets me play both back to compare.
12:12Okay. Uh, let's let's see. Oh, okay.
12:14I need to actually load load an m p three. I think so. I think there was a demo.
12:18I think we gave it some sort of m p three.
12:22Oh. It's twinkle twinkle little sour. Okay.
12:28Alright. Nice. I'm assuming we said that in the system prompt.
12:31Anyway, alright. Load input, analyzing pitch. Alright.
12:33So I I don't know anything about pitch or music notes or anything like that, So I'm just going to judge this based on the interactivity and how well it visually shows this. So okay. Play original.
12:45Okay. Play MIDI. Alright.
12:46You can see alright. Pretty pretty cool. Yeah.
12:48I I'm sure people watching this know what's going on. I have no idea, but it looks like these are piano notes. Somebody please tell me what's going on in the comments.
12:55I know someone's watching that understands this, but I guess we can increase the sensitivity. I don't think that did anything. BPM.
13:02Okay. It's pretty cool. Pretty cool visualizer.
13:04Next up, we have this one. It looks very similar. Load input, very similar, actually.
13:09Jeez. Okay. Uploading audio.
13:11Advanced detection settings. Notes. Okay.
13:14Original. This is very similar look. MIDI.
13:17Okay. It kinda just stops. Okay.
13:18This one's a little bit more broken. Very similar look to the last one. We have this one as well.
13:24Okay. Can you just upload? Can you figure out how to upload this audio file?
13:28One sec. Alright. I can't find the file.
13:30So I'm just gonna say didn't do a good job because these ones were able to load the input. So this one loses. Last one.
13:35Green background again. The new purple gradient looks great, and so we have the input at three. That's cool.
13:42Okay. It doesn't highlight the detected notes, which I don't love. I liked how it did in those last builds, but, uh, yeah.
13:47Okay. I mean, all of these are relatively straightforward. I'm gonna guess this is five six soul.
13:51This has five six soul written all over it. Look at that. This is this is absolutely gonna become the new purple gradient.
13:56Okay. So that's five six soul and then this one is last. Kimi k three.
14:01Darn it. You hate to see it. And then we got what?
14:04Opus? Nope. Kimmy k two seven Opus.
14:06Like, these don't look almost exactly the same. Alright. Might have to move away from that test.
14:10But next up, we have a fun one. Build a two d fighting game. Here is the prompt.
14:15Build me a playable two d fighting game in the browser. Thanks Street Fighter. Distilled.
14:18Two fighters. A health bar for each. Real moves and combos, then we have a bunch of requirements.
14:22Again, prompts are all in the description below. We'll look at this first one. Okay.
14:27Already some text not aligned properly, but okay. We select characters. That's pretty cool.
14:31So player one and then computer. Okay. Two player.
14:35Oh, interesting. Okay. Let's go against the computer.
14:37Volt versus Blaze fight. Okay. I don't know how to punch.
14:40Jump, s crouch, v block. Peter's way better than me. Okay.
14:44I just got destroyed. I still need to figure this out. Okay.
14:47Can I pause this? My goodness. Okay.
14:50W jump, s crouch, f fight, g heavy, b block. Alright.
14:55Wait. We're gonna have to move through this quick, but Okay. Jump, f fight, okay.
14:59Heavy. Yeah. Sounds good.
15:01Pretty fun. Pretty fun. Okay.
15:02Next one. Player versus CPU. Nice.
15:04We got some nice little sound effects. Alright. I like this way better.
15:07S crouch j light k heavy. K. This is kinda weird.
15:11Jeez. Not a gamer. I am not a gamer.
15:13Alright. One more time. Kenji jump s crouch j light k heavy.
15:18I just wanna make contact one time here. I can't even move. Okay.
15:21Jump. Oh, there we go. There we go.
15:22Alright. This one's a little weird. It's like Rock'em Sock'em Robots kind of.
15:25Okay. Number three. This looks like another five six.
15:27I bet this is a five six generation again. Just the font of this background. Okay.
15:32Uh, weird characters. Really weird characters. Soul.
15:35Oh, I wonder what model this is. Gotta be kidding me. Okay.
15:39Rook. Can't even select these. Choose fighter.
15:42I can't even select this. This is crazy. Oh wait.
15:44Okay. That's confusing. Okay.
15:45P one CPU. Alright. That's a confusing thing.
15:47Okay. Alright. Don't love this.
15:49Don't love this at all. Way slower. Okay.
15:52Don't love that one at all. Alright. And then this one.
15:54Alright. Geez. What is going on?
15:56These are all over the place. Player one select. Jeez, look at these icons.
16:01This is terrible. Better be Kimmy k two seven. Okay.
16:03Can't even select this. How do I Player one select, enter? Nope.
16:08This is bad. Okay. That's terrible.
16:09Alright. That's the worst one. 27.
16:11And then I would say this one was second to last, Soul. Look at that, dead giveaway. Um, this one was this one was good.
16:18And then this one Yeah. This one was definitely the most realistic. So I'm gonna say second Kimi k three.
16:23Alright. You go. And then Opus four eight.
16:25Alright. Alright. Fair enough.
16:25Fair enough. Kimi k three, pretty good. I mean, you you can just see the jump between like, first of all, it's better than Soul here.
16:31And second of all, look at the jump between two seven and three. I mean, it's freaking night and day. Like, that is a crazy upgrade.
16:37Okay. Next up, I wanted to do some combination of a landing page and three d. So we're building this luxury watch, landing page, photorealistic rotatable three d watch of the hero that looks like it's shipped from a top brand's launch team.
16:50So we're gonna take a look at all of these. Okay. This yeah.
16:53This is done now. First up, oh, very good three d. Wow.
16:57That's pretty incredible. The the alignment could be fixed. The I actually like this font.
17:02Yeah. Yeah. The actual web design could be use a little bit of work, but the I mean, the three d work is pretty freaking incredible.
17:08Jeez. I've done some, like, generations with Fable, which I know I promised that we would do a couple of those. I don't know if we're gonna get to them, but this is absolutely Fable level.
17:17I don't know if this is k three. I don't know if this is k three, but this is pretty darn good. Okay.
17:20Next up. Jesus Christ. Oh, okay.
17:23Yeah. Not bad. I mean, the the fact that it's rotating like this is very weird.
17:27I do like the dynamic lighting though. That looks good. It doesn't really look like a watch lane, rest of the landing page.
17:33I can actually zoom in. Alright. That's that one.
17:35Next one, kinda weird. K. It's not really rotating properly.
17:40It does look good if it's rotated properly. It looks a little low res too for some reason. Again, these models have the the dynamic lighting down, which is pretty impressive.
17:48Wasn't the case a few months ago. Alright. Not bad.
17:51This one oh, very nice. Very nice. Okay.
17:53Yeah. I actually like this because oh, nice. Okay.
17:56Kind of a weird hideaway, but yeah. Alright. This one it's between this one and this one.
18:00Kinda like this one better even though the placement's wrong. This one looks a little more realistic. Fable oh, I forgot I did Fable five here.
18:07That's one. Uh, okay. I guess I did have a couple of Fable fives in here.
18:10Alright. Uh, Fable five. Well, there well, there you go.
18:12I said this looks a lot like Fable five. It is Fable five. Is this K3?
18:16Opus four eight. Uh, interesting. Okay.
18:18This has to be K3. Oh, 56 Soul and then K three. Interesting.
18:22Interesting. K three didn't do a good job of the watch on the watch test. Hate to see it.
18:26Then I forgot to do Kimi k two seven here, but it probably would have been terrible. I'm realizing too I forgot to include GLM five two in a couple of these. My bad.
18:33My bad. These next builds, when we get into knowledge work, I'll make sure we include GLM five two. Uh, I'm all over the place, but at least we can see how it compares to the frontiers for a lot of these tests.
18:43Alright. Last up, Counter Strike two demo. Counter Strike two demo, you might already be familiar with this one.
18:48I have ran it a couple times with Fable and five six. So if you've seen those videos, you will recognize Fable and five six generations, but we're still going to compare it to k three and see how it does.
18:58This is the prompt, build me a visually stunning genuinely playable browser clone of Counter Strike two. And we just go on to name all of these. But let's take a look at first one.
19:07Sensitivity. Okay. That's a nice touch.
19:10Volume. That's new. So this might be k three.
19:14Woah. Woah. Nice.
19:15Take damage. That's okay. Ran out.
19:17I don't know how to reload. How do I reload? Okay.
19:18R's reload. Wow. This is very good.
19:21Nice. Nice. Yeah.
19:22Okay. And then I think you've seen the rest of these. This oh, okay.
19:25I think this is and now I can't even remember. That might be five six soul. This was Fable's version, and then this I'm assume oh, no.
19:33This was 56Soul. This is 56Soul. There it is.
19:35So okay. This one wasn't bad. Was this this GLM?
19:38Okay. K3. Okay.
19:39Okay. Opus48. Alright.
19:41I had never seen Opus48. Opus4Eight's one was really good. I thought this was Kimmy k three.
19:45I still think that Kimmy k three okay. Let's actually play this. Give Kimmy k three a chance here.
19:50Okay. Terrorist, deploy. Yeah.
19:51I mean, the countdown's a little long. W is not going forward either. That's weird.
19:56So s is go forward. K. That's a weird that's a weird weird weapon.
20:01You're a little slow too. I am very confused on how to go forward. I'm pressing w now.
20:05It's going left. I'm pressing s. It's going right now.
20:09Alright. There we go. Okay.
20:10This is weird. This is weird controls. So not great from Kimi k three.
20:14I would definitely say it's third and then I'm sorry. Fable is definitely number number one still, 56, number 2, Opus number 3, KimiK3, number 4.
20:23Darn, you hate to see it. I thought Kimmy would be doing a better job. The graphics look great.
20:27The actual shooting, not great. The physics and controls is not great. Would have liked to see this tested on GLM 5.2.
20:33My apologies yet again. I'm just trying to rip through these way too fast. Next time, I will make sure we don't miss these, but that is counter strike two.
20:41Alright. So that concludes coding. And I'm gonna be honest, Kimi, at least for these tests, k three, uh, do I dare I say a little underwhelming?
20:49Still a huge jump from k two seven, and it's still right there. It's still it's still competing with the frontier models. I'll give it that.
20:56I'll give it that. Next, what we're gonna do is put it to the test with some knowledge work, and I use knowledge work lightly. I didn't wanna do we've been doing this in the past where I've been doing copywriting and LinkedIn posts and stuff like that.
21:06It just I didn't wanna do that. So this time, we're going to do kind of a blend of knowledge work and coding. We'll have some motion graphics designs, some creative writing, some landing page writing, and like I teased in the intro, a Quentin Tarantino trailer.
21:18Those builds are just wrapping up as well. By the way, I've been paying API costs this whole time. I've probably spent another $25 on top of my subscription.
21:27So I'm at about $65.70 bucks on these three generations today. So let's hit refresh and take a look at some of these knowledge work outputs.
21:34First up, we have an Isaacson style biography intro. So I'm asking it to write an opening paragraph of a biography of myself in the style of Walter Isaacson. If you're not familiar, Walter Isaacson wrote the recent Elon Musk bio.
21:46Of course, the famous Steve Jobs bio. He is an all time writer. And so we're gonna see how these models compare.
21:52We're gonna quickly breeze through these because this is a lot of writing, but let's take a look. Ordinary morning in 2022, freelance marketer in New York opened a little chatbot called GPT three and found himself, as he later recalled, laughing at its hallucinations. So by the way too, I didn't give it any information about me.
22:07All he said was Pat Simmons. So it had to go online, research me, find, I don't know, videos or my website, whatever, and actually put this together. So I mean, fact that it even found this is pretty impressive.
22:17And then, yeah, it's already citing its sources. He was an unlikely guy to the new age, and he said so himself. No technical background, no formal training, and the arcane arts of machine learning.
22:27Nice. Nice. Okay.
22:28Number two. Uh, that wasn't bad. Like like I said, writing is it's so subjective that it is tough, but let's let's just see.
22:35I wanted to try this new kind of unique test here. So next one. Pat Simmons filters the future through the lens of the nontechnical operator.
22:42He does not lecture about algorithms. Instead, he sets two rival models to the same build and lets the artifacts speak for themselves. So this one actually probably found a YouTube video and pulled that transcript to know about that.
22:55The channel stated Charter promises nontechnical AI tools and workflows that can turn a knowledge worker into the AI go to person on a team, and Simmons delivers okay.
23:05This feels more like a, I don't know, Forbes article or something. Okay. Number three, Pat Simmons does not build the machines he judges.
23:12Yeah. No kidding. And he has made that limitation his authority rather than his apology.
23:17Okay. Little AI sounding there, but how many people are actually building machines too? Kind of a weird way to say that.
23:23Second sentence. He came to artificial intelligence sideways out of marketing rather than any laboratory. It's like, okay.
23:28A freelancer by his own account, what followed bore the stamp of a man who trusted his own eyes over the promotional literature. Doesn't seem like this one really understands AI careers, but okay. In a season when artificial intelligence was being sold to the public through leaderboard scores and triumphant benchmarks, Pat Summon's built a small contrarian stage for the opposite idea.
23:49Working from The United States with the unhurried manner of a practitioner rather than a prophet is actually not bad. I like that one. Um, I'm gonna say that one so far is number one.
23:59This one, just annoying copy. Didn't love it. Uh, let's look at the last one, then I will determine.
24:04On 07/10/2026, Pat Simmons opened a forty one minute YouTube review by setting three new AI models to work, and then he kept going. 10 builds.
24:13Okay. So we just looked at one YouTube transcript and wrote this, and then tried to make it sound like a story. Good formatting, though.
24:19That looks cool. Alright. I'm gonna say this one was just annoying, so I'm gonna say it's last.
24:23This one, because it's more of like a biography, I would say number two, three. This one, number four. Alright.
24:29Let's see. Which one what's what's number one? Okay.
24:32GLM5Point2. Interesting. Number two, K3.
24:36Look at that. There we go. Number 3, KimiK 27.
24:39Interesting. Number 456Soul. And then number 5, Opus.
24:43Interesting. This, by the way, is an interesting finding. You can see how good these open source models are are at writing.
24:48GLM five two, which I've done quite a bit of writing with, is actually, yes, very impressive. Uh, this is not just an outlier. And then we have the other open source models coming in the top three, and then the frontier models four eight and five six coming in last.
25:00So this is actually a really good insight here that's worth calling out because coding, even though these are the flashier demos, if you're not a software developer, those mean a lot less to you. Most of us are just in these kind of day to day things where we need a writing partner, where we need something analyzing documents, where we need help just getting through basic knowledge work tasks in our day to day.
25:18Interestingly, all of the open source models did better. It's something that I've seen time and time again, where it really is worth it if you haven't already looking into migrating even just part of your stack over to an open source model. Rant complete.
25:32Next up, we've got a motion graphic explainer. And this guy here, Tatsula, has been putting Kimi k three to the test today, and some of his generations are really cool.
25:42What he did here was he took their launch video here and just had it recreate SVGs. So basically a video clone of Kimmy's video. And look how good it is.
25:52I mean, it's just so impressive. So I wanna do our own test of this by using the hyperframes skill and generating an m p four with motion graphics that is an explainer of Kimi k three.
26:04So let's take a look at that now. Number one, I'm gonna quickly move through these. Okay.
26:08Some good transitions. Wow. These are these are getting good.
26:10This might be one of the better models, but that's, okay, that's really impressive too. It didn't have any any benchmarks, which makes sense why it just says price and there's all these placeholder elements here.
26:20You can see that these motion graphics are really solid. Alright. That's number one.
26:24Number two, they all have this dark background, which is interesting. K. Field guide to Kimi k three.
26:29Alright. Kind of kind of all over the place. Not bad.
26:37Nice little bar chart.
26:42Very cool. Okay. Yeah.
26:44Not bad. What is Kimmy k three? Nice little transition.
26:47Price. Just all placeholders. Alright.
26:49This one's less complex. Still looks pretty good. I'm gonna say this one so far is number one.
26:54This one's lower the pack. Let's see this one now. Some good transitions.
26:58Alright. Kind of boring charts. Hard to read text.
27:01One model. Okay. Yeah.
27:03That one's towards the last, and then something didn't generate. I'll see if can get that generated. But in the meantime okay.
27:07Let's see what this one is. There we go. K three.
27:11Yep. Darn good at motion graphics. Wow.
27:14I mean, look how kinetic and interactive this is. Looks great.
27:18Okay. Which one didn't work? Okay.
27:21GL2 didn't work for some reason. 56Soul. I would say 56Soul is last.
27:26And then this one, I would say this one's number 2. What is this? Oh, it's 48.
27:31This one's number 3. 27. Yeah.
27:33I would say 27 did a better job than 56Soul on that. And then GLN52.
27:38Sorry. It didn't generate. For our next knowledge work test, we have landing page copywriting, a little bit of landing page design.
27:43This is actually for something that I am launching. I'm calling it AI boot camp. Here is the current website.
27:47The link is in the description below if you are interested, but it's essentially a four week live intensive where each week we get into AI fundamentals, defining the business use case for founders, executives, business owners on implementing AI in their own businesses, actually building that out, shipping, all that kind of stuff.
28:03And I've been spending way too long in this landing page just banging my head against the wall. So I figured, why not go to the models and see if they can do any better? So that is what this next test is.
28:13Here is the prompt. Rewrite and redesign the workshop landing page for Pat Simmons' sharper copy that digs into the avatar. So I I really spent a lot of time in the the prompt here because this is the problem I was having.
28:23I I wanted to dig more into pain point and write this in a way that speaks to that pain point a bit more and isn't too salesy, and we're gonna see how these models did. First up, turn we should be using AI into a working tool your business actually uses. Not bad.
28:37That's not really the focus of but okay. Build one live AI tool or workflow for your company in four weeks, then lead your team's AI push knowing what works. Okay.
28:46That's not that's not terrible. That's not a terrible subheader. That's kind of the point.
28:49And then we have this nice little redesign of the header even though this doesn't make a whole lot of sense in this little graphic here. Alright. Wow.
28:57Okay. So it did kind of redesign this whole thing. Right now, AI is probably costing you twice.
29:01K. Digging into the problem. Nice.
29:03Okay. This actually these are these are actually good problem statements. Alright.
29:06The software bill keeps climbing. The busy work is still there. You can feel the gap widening.
29:11AI is on every agenda. Become the person in the room who actually knows. K.
29:15Don't love this design. Spot the right problem. Build with your eyes open.
29:19Lead from experience. Okay. But, yeah, this is this is solid.
29:23This is solid. Four weeks later, the debate is over. It is running.
29:26K. Design's not bad. Learn it.
29:28Find the case. Build it. Ship it.
29:30Small group, real deadlines, no place to hide. I like it. Wow.
29:33I already pulled from a testimonial and then just created this as a pull quote. Alright. This one is I'm depressed.
29:39Not sure what it is though. Next one. Build it yourself first, roll it out second, four week live intensive for founders and executives.
29:45K. Yeah. They're just gonna lean into this.
29:46We should be using AI more, which I don't really wanna lean into a whole lot. But, again, nuance that I didn't probably communicate in system prompts, the subhead way too long. You're not behind because you don't believe in AI.
29:56You're behind because you've never actually built anything with it. Okay. And then we have these problems here, not bad, kind of kind of hard to read and follow along.
30:04None of this is a tooling problem, it's a doing problem. Alright. This one's got AI copy written all over it.
30:10Don't love that one. Okay. Stop renting your operations, Start owning them.
30:14I think they're talking about software, building software. A four week live intensive where founders and executives build AI systems their business actually needs.
30:22Not great. A small cohort of operators who won't let you quit. Em dash dead giveaway.
30:27You're paying for AI twice. Once in software, team barely uses. Again in the hours lost to busy work.
30:33Okay. That's not that's not bad. Quiet dread that you're falling behind.
30:37Quiet. The AIs love using quiet. Okay.
30:39Nice. We've got a little bit of identity shift. Copywriting here.
30:43The operator who stops guessing and starts deciding, not bad, not bad. I think everything else is the same.
30:48So I'm gonna say this one's number one so far. Three, two, but let's see what the rest of these are. Stop reading about AI.
30:53Start building with it. Just not great. AI copywriting has such a long way to go still.
30:58And then this just really long sub headline, a four week live intensive, you pick the workflow that's eating your team's time, build the system that runs it, yada yada yada, way too long. You know you should be further along with AI by now, not because you're behind the news. You probably read more about AI than anyone on your team.
31:14It's so long. You're paying for AI seats. Okay.
31:16Leaning into the problem, not bad. We'll figure it out. AI next quarter is is a decision too late.
31:21This is just so, ugh, so icky. It just feels like you're just being talked down to.
31:26Become the leader who actually knows. Okay. I don't love this.
31:29The fix isn't more research, just one real build. Way too long copy. Okay.
31:33And I think everything else is the same. That one's the last so far. And then what we got?
31:37Okay. Way too long of a headline. In four weeks, you'll be you'll have built the thing you keep saying your company needs.
31:43That's a mouthful. Way too long of a subhead. Don't love this design here.
31:47You've had this meeting. You didn't get where you are by being slow. You've made harder calls than this one.
31:52This one keeps sliding. You approve the budget for AI. It's kind of going like old school copywriters.
31:58It feels like a direct mail piece almost. You bought the seats. Most of them haven't.
32:01Okay. This is just all over the place. Gosh.
32:04More icky copywriting. You already know the other kind of leader.
32:08You don't need to become an engineer. That's not a bad point. One month.
32:10Start to ship. You won't be the only one building. Okay.
32:13Yada yada. Alright. That one, I'll say that one's fourth.
32:16Alright. Big reveal. Fifth?
32:18Oh, no. KimmyK 3.
32:21You hate to see it. That was really quite bad though, if we're being honest. What's number 4?
32:25Opus48. Interesting. Number 3, GLM52.
32:29Number two is Soul really gonna be on the top here? KimmyK 27 is number two? Wow.
32:35And then Soul's number one. I don't know. This could just be an isolated test, but the copywriting was quite terrible.
32:41Anyway, that is workshop landing page. Last test here, and it's a fun one. Quentin Tarantino trailer.
32:47Alright. So for this one, I really wanted to put the model's creative thinking to the test, and the prompt is write and fully shot list a thirty second trailer for a Quentin Tarantino film as a complete ready to generate shot plan for an AI text to video model. So the idea is not only writing a script for a thirty second trailer, but also the complete shot list that we then gave to Sea Dance two to generate.
33:08So we're gonna see how all these look, and then it looks like it also at least a couple of them made actual websites too. I'm really liking what I'm seeing so far. So let's take a look at this one, the last slice.
33:25You're gonna eat that?
33:30It's the last slice.
33:35I know what it is. You boys want a refill? We're deciding some.
33:41Pie gets cold while you decide.
33:50We'll split.
33:53Okay. Not really sure what's going on there. I don't know if it's the video model or the actual script or shot list.
33:58Just random characters kept showing up. Great shot in the opening here though. Nice establishing shot, but everything else kind of all over the place.
34:05I'm not gonna I'm not gonna play all of these, but I'm just gonna do this separately and then I'll be sure to add these to the actual post so you can watch each of these if you are really curious. CDN's, by the way, is just an incredible video model. Look at this.
34:16It just looks so realistic, But that is not the test.
34:21These are so absurd. Alright. This one, I'm just gonna say is the best one.
34:24Which one was it? Oh, GLM52. There we go.
34:26This, very confusing. I'm gonna say fourth. This one just because the shots prompted.
34:31I say number two. This one, I'll give number 3 because of that establishing shot, and then this number 5. Let's see.
34:3748. Okay. 56 soul number 2.
34:39The K27Number4. K3Number3. Alright.
34:42Not a great test. I'm just trying to get creative with it, see if we can push these models and do something a little bit different. But, uh, there we have it.
34:48So with that, let's do one final tally of where Kimmy k three landed across the board. So here is what we're looking at. On the coding side, it came in at first, fourth, third, second, fourth.
34:59The interactive gadget, that three d rendering was pretty impressive. We kicked it off, and I thought we're gonna see some better outputs from k three, but these are truly so subjective. And then a knowledge work, it did a little bit better.
35:10Second, first, and then fifth, third. So I wasn't totally blown away like I maybe was with Fable or five six, but that's also not the point with these open source models.
35:20If we looked at some Kimi k two seven generations for example, you can see how much of a step up it is from two seven to now three. And I mentioned this too, but for the most part, you're not doing crazy hard coding tasks. A lot of this is just trying to get through your day.
35:33And for an open source model that is much cheaper than any of the Frontiers, Kimi k three seems like a good selection. But, uh, curious what you think. Appreciate the patience.
35:42I know this is a long one. This one kinda got out of control. I'm really trying to work on reining in these model release demos.
35:48Sometimes I get a little crazy with it. Continuing to just try to test the creativity, the kind of day to day tasks especially to see how well these truly do.
35:55So appreciate you sticking with me. As always, if you have any ideas for future demos to run-in any of these model release videos, please let me know, and I'll see you in the next one.
The Hook

The bait, then the rug-pull.

Moonshot AI's Kimi K3 landed with benchmark charts claiming it goes toe-to-toe with the frontier labs — so this video skips the leaderboards and fans the same ten build prompts out across Kimi K3, its predecessor, and four rival models, ranking every output blind before revealing which model made what.

Frameworks

Named ideas worth stealing.

01:00concept

Agent fan-out testing

Run the identical prompt across six terminal windows wired to different backing models via Claude Code — some accessed through OpenRouter, some through direct subscriptions, one through a separate CLI login — so every model builds the same spec simultaneously and can be blind-ranked before reveal.

Steal forany multi-model bakeoff or internal model-evaluation pipeline
CTA Breakdown

How they asked for the click.

VERBAL ASK
08:00subscribe
“please like this video, subscribe to the channel if you haven't already”

Dropped casually mid-video during a Kimi rate-limit wait, framed as helping subsidize the host's AI subscription costs rather than a hard sales pitch.

MENTIONED ON CAMERA
00:30toolAgent fan-out skill (GitHub) ↗
27:40productAI Boot Camp waitlist ↗
36:06linksimmonsbench.com/kimi-k3 (full generations + prompts) ↗
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Storyboard

Visual structure at a glance.

cold open
hookcold open00:00
3D gadget test
value3D gadget test06:41
weather app test
valueweather app test09:05
MP3-to-MIDI test
valueMP3-to-MIDI test11:39
fighting game test
valuefighting game test14:10
watch page test
valuewatch page test16:37
Counter-Strike test
valueCounter-Strike test18:43
biography test
valuebiography test21:34
motion graphics test
valuemotion graphics test25:31
landing page test
ctalanding page test27:40
Tarantino trailer
valueTarantino trailer32:45
final tally
closefinal tally34:48
Frame Gallery

Visual moments.

cold open00:00Frame at 00:34 from Kimi K3 Is Here! (Better Than Opus 4.8?)00:34Frame at 01:00 from Kimi K3 Is Here! (Better Than Opus 4.8?)01:00Frame at 01:26 from Kimi K3 Is Here! (Better Than Opus 4.8?)01:26Frame at 02:01 from Kimi K3 Is Here! (Better Than Opus 4.8?)02:01Frame at 02:28 from Kimi K3 Is Here! (Better Than Opus 4.8?)02:28
cold open00:00Frame at 00:34 from Kimi K3 Is Here! (Better Than Opus 4.8?)00:34Frame at 01:00 from Kimi K3 Is Here! (Better Than Opus 4.8?)01:00Frame at 01:26 from Kimi K3 Is Here! (Better Than Opus 4.8?)01:26Frame at 02:01 from Kimi K3 Is Here! (Better Than Opus 4.8?)02:01Frame at 02:28 from Kimi K3 Is Here! (Better Than Opus 4.8?)02:28Frame at 02:55 from Kimi K3 Is Here! (Better Than Opus 4.8?)02:55Frame at 03:23 from Kimi K3 Is Here! (Better Than Opus 4.8?)03:23Frame at 04:03 from Kimi K3 Is Here! (Better Than Opus 4.8?)04:03Frame at 04:11 from Kimi K3 Is Here! (Better Than Opus 4.8?)04:11Frame at 04:44 from Kimi K3 Is Here! (Better Than Opus 4.8?)04:44Frame at 05:11 from Kimi K3 Is Here! (Better Than Opus 4.8?)05:11Frame at 05:38 from Kimi K3 Is Here! (Better Than Opus 4.8?)05:38Frame at 06:05 from Kimi K3 Is Here! (Better Than Opus 4.8?)06:05Frame at 06:38 from Kimi K3 Is Here! (Better Than Opus 4.8?)06:383D gadget test06:54Frame at 07:23 from Kimi K3 Is Here! (Better Than Opus 4.8?)07:23Frame at 07:47 from Kimi K3 Is Here! (Better Than Opus 4.8?)07:47Frame at 08:28 from Kimi K3 Is Here! (Better Than Opus 4.8?)08:28Frame at 08:45 from Kimi K3 Is Here! (Better Than Opus 4.8?)08:45weather app test09:05Frame at 09:44 from Kimi K3 Is Here! (Better Than Opus 4.8?)09:44Frame at 10:00 from Kimi K3 Is Here! (Better Than Opus 4.8?)10:00Frame at 10:36 from Kimi K3 Is Here! (Better Than Opus 4.8?)10:36Frame at 11:05 from Kimi K3 Is Here! (Better Than Opus 4.8?)11:05Frame at 11:19 from Kimi K3 Is Here! (Better Than Opus 4.8?)11:19MP3-to-MIDI test11:57Frame at 12:13 from Kimi K3 Is Here! (Better Than Opus 4.8?)12:13Frame at 13:04 from Kimi K3 Is Here! (Better Than Opus 4.8?)13:04Frame at 13:05 from Kimi K3 Is Here! (Better Than Opus 4.8?)13:05Frame at 13:33 from Kimi K3 Is Here! (Better Than Opus 4.8?)13:33Frame at 14:01 from Kimi K3 Is Here! (Better Than Opus 4.8?)14:01fighting game test14:39Frame at 15:01 from Kimi K3 Is Here! (Better Than Opus 4.8?)15:01Frame at 15:24 from Kimi K3 Is Here! (Better Than Opus 4.8?)15:24Frame at 15:53 from Kimi K3 Is Here! (Better Than Opus 4.8?)15:53Frame at 16:17 from Kimi K3 Is Here! (Better Than Opus 4.8?)16:17watch page test16:44Frame at 17:20 from Kimi K3 Is Here! (Better Than Opus 4.8?)17:20Frame at 17:49 from Kimi K3 Is Here! (Better Than Opus 4.8?)17:49Frame at 18:03 from Kimi K3 Is Here! (Better Than Opus 4.8?)18:03Counter-Strike test18:35Frame at 19:05 from Kimi K3 Is Here! (Better Than Opus 4.8?)19:05Frame at 19:24 from Kimi K3 Is Here! (Better Than Opus 4.8?)19:24Frame at 20:12 from Kimi K3 Is Here! (Better Than Opus 4.8?)20:12Frame at 20:21 from Kimi K3 Is Here! (Better Than Opus 4.8?)20:21Frame at 20:58 from Kimi K3 Is Here! (Better Than Opus 4.8?)20:58Frame at 21:16 from Kimi K3 Is Here! (Better Than Opus 4.8?)21:16biography test21:54Frame at 22:14 from Kimi K3 Is Here! (Better Than Opus 4.8?)22:14Frame at 22:47 from Kimi K3 Is Here! (Better Than Opus 4.8?)22:47Frame at 23:07 from Kimi K3 Is Here! (Better Than Opus 4.8?)23:07Frame at 23:35 from Kimi K3 Is Here! (Better Than Opus 4.8?)23:35Frame at 24:02 from Kimi K3 Is Here! (Better Than Opus 4.8?)24:02Frame at 24:35 from Kimi K3 Is Here! (Better Than Opus 4.8?)24:35Frame at 24:49 from Kimi K3 Is Here! (Better Than Opus 4.8?)24:49motion graphics test25:36Frame at 25:54 from Kimi K3 Is Here! (Better Than Opus 4.8?)25:54Frame at 26:13 from Kimi K3 Is Here! (Better Than Opus 4.8?)26:13Frame at 26:42 from Kimi K3 Is Here! (Better Than Opus 4.8?)26:42Frame at 27:04 from Kimi K3 Is Here! (Better Than Opus 4.8?)27:04Frame at 27:33 from Kimi K3 Is Here! (Better Than Opus 4.8?)27:33landing page test28:03Frame at 28:32 from Kimi K3 Is Here! (Better Than Opus 4.8?)28:32Frame at 29:13 from Kimi K3 Is Here! (Better Than Opus 4.8?)29:13Frame at 29:38 from Kimi K3 Is Here! (Better Than Opus 4.8?)29:38Frame at 30:07 from Kimi K3 Is Here! (Better Than Opus 4.8?)30:07Frame at 30:26 from Kimi K3 Is Here! (Better Than Opus 4.8?)30:26Frame at 30:48 from Kimi K3 Is Here! (Better Than Opus 4.8?)30:48Frame at 31:26 from Kimi K3 Is Here! (Better Than Opus 4.8?)31:26Frame at 31:36 from Kimi K3 Is Here! (Better Than Opus 4.8?)31:36Frame at 32:07 from Kimi K3 Is Here! (Better Than Opus 4.8?)32:07Frame at 32:40 from Kimi K3 Is Here! (Better Than Opus 4.8?)32:40Tarantino trailer33:16Frame at 33:37 from Kimi K3 Is Here! (Better Than Opus 4.8?)33:37Frame at 33:56 from Kimi K3 Is Here! (Better Than Opus 4.8?)33:56Frame at 34:19 from Kimi K3 Is Here! (Better Than Opus 4.8?)34:19final tally34:48Frame at 35:29 from Kimi K3 Is Here! (Better Than Opus 4.8?)35:29Frame at 35:52 from Kimi K3 Is Here! (Better Than Opus 4.8?)35:52
Watch next

More from this channel + related breakdowns.

40:03
Pat Simmons · Review

GPT-5.6 Sol: No-Hype Full Review & Testing

A blind, four-way bake-off — GPT-5.6 Sol against Fable, Opus 4.8, and GPT-5.5 — across ten builds and knowledge-work tasks, scored one task at a time without knowing which model made what.

July 10th
19:10
Pat Simmons · Tutorial

Clone a $1.2B App in 19 Minutes

An 8-step agentic pipeline that takes you from naive AI slop to a pixel-near Linear replica, deployed to Vercel with an MCP server, in under 20 minutes.

June 8th
Chat about this