Kimi K3.1 Pro Drops Overnight: 2.8T Parameters, Terminal Score Just 0.7 Behind Opus
Today, in the early hours, Kimi K3.1 Pro was released.
In this demo video, a 3D printer is printing parts. Assemble it, plug in power, and the light turns on.
The model for this lamp was made by Kimi.
2.8 trillion total parameters, 1 million token context. Token can be roughly understood as a small unit used when AI reads and writes content. The latter number refers to how much material it can process at once.
The new model is now online, and the complete weights are planned to be released before November 4.
Leave the lamp for later, first look at the game it made this time — with a map, combat, and a two-phase boss.
5 regions, two-phase boss, the adventure game is done
This 3D action-adventure game has a cartoonish art style and uses Godot 4.7.2.
Godot, like Unity and Unreal Engine, is a game engine. Simply put, it's a tool that brings characters, scenes, controls, and gameplay together to make a game.
The entire project has 5 explorable areas, with villages, terrain, roads, water, and vegetation all set up.
The map is done, but there needs to be gameplay.
It has third-person combat, and the boss fight is split into two phases. The character can gather resources, the map has day-night cycles, and there is fast travel between campfires.
Damn, AI game-making has reached this level?
Looking at how it was made is quite practical. Plants, villages, characters, and animations used ready-made assets from asset libraries like Quaternius, Kenney, and KayKit, and Kimi integrated them into the game.
This is very close to how actual projects are done — use suitable assets and spend time making them all work together.
36 million triangles, a dragon flies into the UE heavenly palace
Moving to Unreal Engine 5.8, also known as UE, the art style immediately changes.
An Eastern heavenly palace, grand halls, colonnades, a dragon approaching from afar.
The protagonist is made with MetaHuman, a digital human toolset in UE. The dragon was procedurally modeled in Blender, which is commonly used for 3D modeling and animation.
The source model of this scene alone has 36 million triangles. The 3D model you see is composed of these small faces.
The scene also uses UE's Nanite and Lumen, the former handling the display of complex models and the latter handling lighting.
Next, the perspective goes to the sky.
A swordsman rides a sword through the sea of clouds, flies over floating islands, and enters an ancient city above the clouds. This is an oriental-style RPG preview video made with UE 5.8. RPG stands for role-playing game.
The concept design and some asset inspiration for this segment come from Xiaohongshu creator @-Duduu-. It also uses CC0 3D scan assets from museums, which are open assets that can be freely used.
Find references, pick assets, set up scenes, then connect actions and camera. Behind a segment, there's a lot of work to stitch together.
The official also listed Unity in the supported game engine list. In these specific demos, we see the Godot adventure game, and the UE heavenly palace and sword-riding preview.
From a robot short film to a lamp that actually lights up
Blender's work goes beyond just building models.
In this robot short film, an old metal workshop, city ruins, rising dust, and evening light are all made. The robot also has its own expressions and actions.
Blender 5.2 and Python are used in production. Python is a programming language, here used to control modeling and animation.
Some assets like props, rocks, textures, and sky are ready-made. Rendering, video output, and sound are also handled afterward.
Models, animation, lighting, sound, all combined into one short film.
The film is ready to deliver. But the lamp at the beginning still needs to be taken out of the screen.
Its design is inspired by Fujian tulou (earthen buildings), divided into 5 parts. The lampshade, base, and internal structure can be viewed individually or see how they assemble.
Then, the camera cuts from the computer screen to the 3D printer, and the parts start printing layer by layer.
Assemble it, light it up. The model that was rotating on screen moments ago can now sit on the desk.
The newly launched 3D Make can create models from text, sketches, or reference images. After creation, you can preview online, continue editing, download for 3D printing, and export video and live photos.
Furniture and mechanical parts are also in the showcase. K3.1 Pro can write scripts to call Blender and CAD software to modify models. CAD is commonly used for engineering drawings and part design, where dimensions must match.
For example, making a cabinet: if you change the cabinet width, the partitions and openings adjust accordingly. This is called parametric modeling.
36 calculation errors, it went back and fixed the simulator
After all the beautiful finished products, here is one that didn't go smoothly at first.
Kimi built a walking simulator based on the open-source human anatomy model human-atlas, letting a simulated person in the computer learn to walk.
The earliest version couldn't stand for long before falling.
It continued training and fixing, but for a while, the arms just wouldn't adjust correctly.
The training has a reward system—do the right thing and score points—to guide it to learn walking. But adjusting this system alone didn't solve the problem.
It went back to inspect the simulator and finally found 36 errors, all in the calculations of how shoulder muscles drive arm movement.
After fixing these and retraining, the simulated person could finally alternate left and right steps, walking continuously for 11.2 seconds. The entire experiment involved 167 tasks and 420 reviews.
People often talk about Agent—it can be understood as an AI that calls tools and works step by step. Code is written, then run, results are checked, and if problems arise, go back and fix.
This back-and-forth is a completely different way of working compared to handing in a piece of code and being done.
Terminal 65.7%, CAD modeling 92.7%
We've seen official demos, now look at test scores.
Terminal-Bench 4.0 tests whether an AI can complete engineering tasks in a terminal. A terminal is the window where you type commands and run programs.
In the official aggregated scores, Kimi K3.1 Pro is 65.7%, Claude Opus 5.5 is 66.4%, GPT-6 Astra is 58.2%.
In this category, Kimi is only 0.7 percentage points behind Opus.
For 3D modeling, the test changes.
BenchCAD gives an AI several images from different angles and asks it to write CAD code to build the model.
With tools allowed, Kimi scores 92.7%, Opus 96.2%, Astra 95.9%. In the chart, "with tools" means tools are allowed.
Kimi also did an internal comparison specifically for multi-step long tasks.
This test, called Kimi Code Bench 3.0 Long-horizon, mainly looks at whether an AI can continuously do many steps, call tools, and complete engineering tasks from start to finish.
K3 scored 55.95%, K3.1 Pro reached 61.2%, an improvement of 5.25 percentage points.
K3.1 Pro is a continued training on top of K3, with 1 million token context, which the previous generation also had.
Nvidia financial model, formulas and predictions delivered together
Beyond games and animation, the official also delivered a few office documents: an Nvidia financial valuation model, a CATL investment committee presentation, and a quantum computing industry report.
The investment committee is where people discuss whether to invest. The presentation must clearly state the judgment and evidence.
The financial model's key is in the spreadsheets.
Multiple sheets are interconnected with formulas that can be clicked open. Change a number up front, and the results after it change accordingly.
The balance sheet below shows actual data, prediction columns, and formulas.
The research report connects another thing: seeing a claim, following the citation, you can find the original source it used.
Spreadsheets that continue calculations, reports that trace back citations. People who make materials every day would know how much back-and-forth searching this saves.
On OfficeQA Pro v2, K3.1 Pro's official score is 64.4. This test requires the AI to find information from real financial documents, compare across several files, and complete calculations.
Final thoughts
In the past, a person might have a world in their head, but in the end only a piece of text remains. Not necessarily the story they most wanted to write, because modeling, programming, and animation all had to be learned from scratch.
If these tasks can be delegated more to AI, perhaps the order of choices can be reversed.
First ask: should this world be read, seen, or walked into?
Thank you for reading my article. If you like it, feel free to give a like, save, and share—three in a row 🙂 If you want to be notified first, you can also give me a star⭐～Thanks for reading my article.
Open-source knowledge base (real-time updating, discussion group): https://my.feishu.cn/wiki/OhQ8wqntFihcI1kWVDlcNdpznFf