On today's chart 951K views since yesterday

Kimi K3.1 draws Fireship a fine horse, then animates it badly

Jeff Delaney's Code Report weighs China's free, three-trillion-parameter model, the fight over distillation, and a homemade test built on hooves.

Channel:

Fireship Story of October 27, 2026Trari Newsroom

Why this story? One of the most watched videos since yesterday, with 951,180 views.

China just dropped a new frontier model... and nobody can explain how

Things we can learn from the video

Press to jump

0:00An anonymous Chatbot Arena model called Kibitz is revealed as Moonshot AI's Kimi K3.1 Pro.
0:27Kimi K3.1 Pro is a three-trillion-parameter open-weight model with strong reported scores.
0:54Anthropic and OpenAI accuse Chinese companies of distilling their frontier models; the host jokes it is a human right.
2:09Terminal-Bench and SWE Marathon results put Kimi close to closed models such as Opus.
2:46Moonshot removed cybersecurity exploit data on purpose to reduce the model's hacking abilities.
4:23On ThoroughBench, Kimi builds a detailed 3D horse, but animates it with crude mesh warping.

Mystery model, unmasked. An anonymous entry called "Kibitz" turned up on Chatbot Arena, and it turned out to be Kimi K3.1 Pro from Moonshot AI. That reveal opens Fireship's latest Code Report, which drew about 950,000 views in its first nine hours.

The pitch, as Jeff Delaney lays it out: a free, open-weight model with three trillion parameters that, on the Terminal-Bench coding test, comes within a point of Anthropic's Opus. A week earlier the channel was test-driving Google's new model against GPT-6 and Opus. This time the challenger comes from China, and the title asks how.

Part of the answer is the argument hanging over the whole industry. Anthropic and OpenAI have accused Chinese companies of distillation, training their own models on the outputs of frontier ones. Delaney does not hide where his sympathies lie. "Now, I believe distillation is a fundamental human right," he says, before describing how the big labs see it: "Grand Theft Auto-Complete," a practice they have spent millions trying to stop. He also notes a less expected choice: Moonshot deliberately stripped cybersecurity exploit data from its training to blunt the model's hacking skills.

Then the leaderboards give way to ThoroughBench, Delaney's own test of how well a model can build an interactive 3D horse with Three.js. Kimi produces a handsome, detailed animal. Then it has to move. The motion comes from simple mesh warping, and Delaney's verdict is short: "The animations are absolutely terrible." The on-screen caption puts it more bluntly.

The horse has taken over the comments. Many viewers want it made an official benchmark for every new model, and others joke that they follow the channel mainly for the equine updates. The episode closes with a sponsored segment for the design tool Mobbin.

Published on Trari

Where next?

Also in Technology companies and softwareThis week on the channel +4,031 views this weekOliur Online's design cleanup starts with fonts, not more featuresAn older Claude tutorial strips away emojis, purple accents and awkward spacing during a week without new uploads.Channel: Oliur Online · Story of November 11, 2026

A different discovery todayVideo taking off 49x its usual first dayWomen in ITV's Power and Silence say David Walliams pursued them as teenagersTwo of the women were 15 when they met him. Walliams denies all wrongdoing, saying in a statement: "I am not a sexual predator".Channel: ITV News · Story of November 11, 2026

What viewers say

AI companies: trillion parameters.

Me: show me the horse.
@Davidayo_See it on YouTube
I only watch fireship to keep up with the development of horse tinder
@allrevenuegoneSee it on YouTube
You have a suspiciously large collection of buff guy images
@thedavymacSee it on YouTube

7 more comments

We need an oficial horse benchmark score for all AIs
@forealoveSee it on YouTube
Last time i was this early AI was fancy autocomplete
@fff-kd8kgSee it on YouTube
Well you forgot to add "make no mistakes". No wonder the horse moves like that.
@MrbiernotSee it on YouTube
5:14 argon being accurate
@catm382See it on YouTube
2:02 Trust me bro, this trust me bro is the trust me bro that we can trust, bro.
@ikegru4346See it on YouTube
THE TRUST ME BRO BENCHMARKS 😂
@AdamsOyifiSee it on YouTube
The reason the 3D model of the horse looked so real is that Kimi uses Tripo to generate the 3D model and then simply displays it as a .glb file. That’s why it couldn't rig it properly, resulting in a horrible walk animation.
@TheMetaverseGuySee it on YouTube

Public comments from the video's page on YouTube

Or follow us on Telegram

The facts

Channel: Fireship
Subscribers: 4,320,000 ▲ 10,000 this week
Channel views, last 7 days: 7,934,291
Channel views, pace per day: 1,274,568 (on a normal day: 1,130,461)
Video published: October 26, 2026
Video views so far: 951,180 (less than a day online)
Video views, pace per day: 1,710,414

The numbers

A big start for one of tech YouTube's most watched news shows, in line with how fast Fireship's videos usually open.

subscribers

4,320,000

▲ 10,000 in the last 7 days

total views

790,281,819

views per day

1,274,568

typical day: 1,130,461

the video

views

951,180

likes

18,586

comments

935

days online

subscribers: this channel and similar channels

4,480,000

4,320,000

1,360,000

1,190,000

1,190,000

1,150,000

As of October 27, 2026 · Data source: Trari Social News, Trari Traffic Trends and Ranking Platform, Google YouTube

Subscriber, view, like and comment counts are YouTube's. Rates, usual days, rankings, comparisons and categories are Trari's own calculations.

💡 Test your knowledge

Charlie Day, heard in an archive clip in this episode, stars in It's Always Sunny in Philadelphia. Which network first aired the show?