今天凌晨，Kimi K3.1 Pro发布。
这次的演示视频里，3D打印机正在打零件。装起来，接上电，灯亮了。
这盏灯的模型，是Kimi做的。
2.8万亿总参数，100万Token上下文。 Token可以粗略理解成AI读写内容时用的小单位。后面这个数字，说的是它一次能处理多长的材料。
新模型现在已经上线，完整权重计划在11月4日前放出。
灯放后面看，先看看它这次做出来的游戏——有地图，有战斗，还有两阶段Boss。
5个区域、两阶段Boss，冒险游戏做出来了
这款3D动作冒险游戏，画风偏卡通，用的是Godot 4.7.2。
Godot和Unity、Unreal Engine一样，都是游戏引擎。说白了，就是把角色、场景、操作和玩法放在一起，做成游戏的工具。
整个项目有 5个可以探索的区域，村庄、地形、道路、水面和植被都搭好了。
地图做好了，还得有得玩。
它做了第三人称战斗，Boss战分成两个阶段。角色可以采集，地图有昼夜变化，还能在营火之间快速传送。
卧槽，AI做游戏已经到这一步了？
往下看它怎么做的，倒也很实在。植物、村庄、角色和动画，用了Quaternius、Kenney、KayKit等素材库里的现成素材，Kimi再把它们接进游戏里。
这就很接近实际做项目的方式了，合适的素材拿来用，把时间花在怎么让它们一起跑起来。
3600万三角面，巨龙飞进UE天宫
到了Unreal Engine 5.8，也就是常说的UE，画风一下就换了。
东方天宫，大殿、柱廊，一条巨龙从远处靠过来。
主角用MetaHuman制作，这是UE的一套数字人制作工具。巨龙则是在Blender里用程序建模的，Blender就是常用来做3D模型和动画的软件。
光这个场景的源模型，就有 3600万个三角面。你看到的3D模型，就是由这些小面拼起来的。
场景还用了UE的Nanite和Lumen，前者负责处理这些复杂模型的显示，后者负责光照。
接着，视角到了天上。
剑客御剑穿过云海，飞过浮岛，再进入云上的古城。这是一支用UE 5.8做的东方风格RPG预览片，RPG就是角色扮演游戏。
这段的概念设计和部分素材灵感，来自小红书创作者@-Duduu-。里面还用了博物馆公开的CC0三维扫描素材，也就是可以自由使用的那类开放素材。
找参考、挑素材、搭场景，再把动作和镜头串起来。一段片子背后，要接起来的活儿其实不少。
官方也把Unity列进了支持的游戏引擎名单。这批具体演示里，能看到的是Godot的冒险游戏，以及UE里的天宫和御剑预览片。
从机器人短片，到一盏真的能亮的灯
Blender接到的活儿，就不止建个模型了。
这部机器人短片里，旧金属车间、城里的废墟、扬起来的灰尘、傍晚的光，都做了。机器人也有自己的表情和动作。
制作时用了Blender 5.2和Python，Python是一门编程语言，在这里用来控制建模和动画。
道具、石头、贴图和天空用了一部分现成素材。后面的画面渲染、视频输出和声音，也接着处理好了。
模型、动画、灯光、声音，最后合在一整段片子里。
片子做到这里，可以交了。但开头那盏灯，还要从屏幕里拿出来。
它的设计灵感来自福建土楼，一共拆成5个零件。灯罩、底座、里面的结构，都能单独看，也能看它们怎么装回去。
然后，镜头从电脑画面切到了3D打印机，零件开始一层层打出来。
装好，点亮。刚才还在屏幕里转的模型，现在能摆到桌上了。
这次上线的3D Make，可以用文字、草图或者参考图来做模型。做完能在线预览、继续改，下载去3D打印，也能导出视频和实况照片。
家具和机械零件也在展示范围里。K3.1 Pro会写脚本，调用Blender和CAD软件来改模型。CAD常用来画工程图、设计零件，尺寸得对得上。
比如做个柜子，你把柜宽改了，隔板和开口也跟着改，这就叫参数化建模。
36处计算错误，它回头修了模拟器
漂亮的成品看了不少，下面这个，开头就没那么顺了。
Kimi基于开源人体解剖模型human-atlas，搭了一个行走模拟器，让电脑里的模拟人学着走路。
最早的版本，站不了多久就倒。
它接着训练、接着改，可有一阵子，手臂怎么调都不对。
训练里有一套奖励规则，做对了就得分，用来引导它学会走路。但光调这套规则，问题没解决。
它回头查模拟器，最后发现了 36处错误，都出在肩部肌肉怎么带动手臂的计算上。
把这些地方修掉，再训练，模拟人终于能左右脚交替迈步，连续走了11.2秒。整次实验一共做了167项任务、420次复核。
大家常说的Agent，可以理解成会调用工具、一步步干活的AI。代码写完，还要运行、看结果，出了问题再回去改。
这一来一回，和交一段代码就结束，已经是两种工作方式了。
终端65.7%，CAD建模92.7%
前面看的是官方演示，再看看测试成绩。
Terminal-Bench 4.0考的是让AI在终端里完成工程任务。终端就是那个输入命令、运行程序的窗口。
在官方汇总的成绩里，Kimi K3.1 Pro是65.7%，Claude Opus 5.5是66.4%，GPT-6 Astra是58.2%。
这一项，Kimi和Opus只差0.7个百分点。
到了3D建模，考法就变了。
BenchCAD会给AI几张不同角度的图，让它写CAD代码，把模型做出来。
允许调用工具时，Kimi的成绩是92.7%，Opus是96.2%，Astra是95.9%。图里的with tools，说的就是可以用工具。
Kimi还做了一组内部对比，专门看多步骤的长任务。
这项测试叫Kimi Code Bench 3.0 Long-horizon，主要看AI能不能连续做很多步、调用工具，把工程任务从头做到尾。
K3是55.95%，K3.1 Pro到了61.2%，提高了5.25个百分点。
K3.1 Pro是在K3基础上继续训练出来的，100万Token上下文，上一代就有。
英伟达财务模型，公式和预测一起交付
游戏和动画之外，官方还交了几份办公文件：英伟达的财务估值模型、宁德时代的投委会演示文稿，还有量子计算行业报告。
投委会，就是讨论要不要投资的那群人开会。演示文稿得把判断和依据讲清楚。
财务模型这一份，关键在表格里面。
几张表是连着算的，公式也能点开看。你改了前面的一个数字，后面的结果就跟着变。
下面这张资产负债表，能看到实际数据、预测栏和公式。
研究报告则把另一件事接上了：看到一个判断，顺着引用，就能找到它用的原始资料。
表格能接着算，报告能回头查。 天天做材料的人，应该知道这能省多少来回翻找的工夫。
在OfficeQA Pro v2上，K3.1 Pro的官方成绩是64.4。这项测试会让AI从真实财务文件里找资料，对照几份文件，再完成计算。
写在最后
过去，一个人脑子里有个世界，最后却可能只留下一段文字。未必是他最想写故事，只是建模、写程序、做动画，哪样都得从头学。
如果这些活儿能更多地交给AI，选择的顺序也许可以倒过来。
先问：这个世界，应该让人读到，看到，还是走进去？
谢谢你读我的文章。
如果觉得不错，随手点个赞、在看、转发三连吧🙂
如果想第一时间收到推送，也可以给我个星标⭐～谢谢你看我的文章。
开源知识库（实时更新 交流群）：
my.feishu.cn/wiki/OhQ8wq…
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
This morning, Kimi K3.1 Propublish.
In this presentation, 3D printers are playing parts. Load it up, turn on the power, the light's on.
The model of this lamp was made by Kimi.
2.80 trillion total parameter, 1 million Tokencontext. Token can be roughly understood as the unit used to read and write in AI. The latter figure is the amount of material it can handle at once.
The new model is now online and the full weight is scheduled to be released by 4 November.
The lights are in the back. Let's take a look at the game it's made. There's maps, there's fighting, there's two stages of Boss.
Five regions, two stages, Boss.
This 3D action risk-taking game, which draws a windy cartoon, uses Godot 4.7.2.
Godot and Unity and Unreal Engineering are game engines. To put it bluntly, it is to put the role, scene, operation and play together as a game tool.
The whole project has five areas to explore, with villages, terrain, roads, water and vegetation.
The map is ready, and we have to play.
It did a third person's call to battle, and Boss split into two stages. Roles can be collected, maps change day and night, and can be transmitted quickly between camp fires.
God damn it, is this the point of Ai playing?
Looking down at how it works, it's real. Plants, villages, roles and animations, using ready-made assets from asset vaults such as Quternius, Kenney, KayKit and so on, were then taken into the game by Kimi.
It's very close to the way the project is actually done. The right asset is used to spend time on how to get them running together.
36 million triangles, dragons flying into the UE Palace.
When we get to Unreal Engineering 5.8, which is the usual UE, the style changes.
The Orient Sky Palace, the temple, the corridor, a giant dragon coming from afar.
The lead was produced by Methauman, a digital human production tool for UE. The dragon is modelled in a program in Blender, which is the software used to make 3D models and animations.
The source model of this scene alone is 36 million triangles. The 3D model you saw, it was put together by these little faces.
The scenario also uses UE ' s Nanite, which handles the display of these complex models, and Lumen, which handles light.
And then, the perspective went to heaven.
The sword of the swordsmen passed through the clouds, flew over the floating islands and entered the ancient cities of the clouds. It's an Orient-style RPG preview with UE 5.8. RPG is a role-playing game.
The conceptual design and part of asset inspired this section from REDnote creator@-Duduuu. It also uses an open 3-D scan of asset by the museum, i.e. open asset, which is freely available.
Look for reference, pick asset, set the scene, then tie up the action with shot. There's a lot of work to be done behind a film.
The government has also placed Unity on the list of supported games engines. In this demonstration, you can see Godot's adventure game, and the UE skyscraper and sword preview.
From a short robot film to a real light.
There's more than a model for the job that Blinger received.
In this short film, the old metal workshop, the ruins of the city, the dust that rises, the light of the evening is done. The robots have their own faces and movements.
It was produced in the language of Blender 5.2 and Python, a programming language used to control modelling and animation.
Some of the props, stones, stickers and skyscrapers are available in asset. The image, the video output and the sound in the back were also processed.
Models, animations, lights, sound, and finally a whole film.
The film does it here, it's ready. But the first light is to be removed from the screen.
Its design was inspired by the construction of a building, which was demolished into five parts. The lights, the floors, the structures inside them can be seen alone and how they are put back.
shot then cut from the computer image to the 3D printer, and the spare parts began to come out on one level.
Set it up, light it. The model that just turned on the screen can now be on the table.
3D Make, this time online, can be modelled by text, sketch or reference. The online preview, continuous adaptation, download to 3D printing and the export of video and live photographs.
Furniture and mechanical parts are also on display. K3.1 Pro writes scripts and calls the Blender and CAD software to change the model. CAD is often used to draw engineering drawings, design parts, and is appropriate in size.
It's like making a cupboard, you change the width, the partition and the opening. It's called parametric modelling.
Thirty-six arithmetical errors. It went back to repairing the simulator.
It's not that good at the beginning.
Kimi built a walk simulator, based on the open source human anatomy model, hman-atlas, to enable the computer simulators to learn to walk.
The original version, it won't stand long.
It goes on to train, then change, but for a while, the arms can't be right.
There's a reward rule in the training, and it's the right thing to do, and it's the right thing to learn to walk. But the optic rule, the problem is not solved.
It looked back at the simulator, and it found 36 errors, all of which were in the calculation of how the shoulder muscles drove the arm.
Fixing these places, retraining them, the simulators can finally walk around their feet and walk for 11.2 seconds. A total of 167 missions and 420 reviews were conducted throughout the experiment.
Agent, as you often say, can be understood as an AI that calls tools and works step by step. When the code is finished, you have to run, see the results and go back and change it.
This time, and the handing over of a code, is already two ways of working.
Terminal 65.7%, CAD modelling 92.7%
It's an official presentation. Let's see the test results.
Terminal-Bench 4.0 was given to allow AI to complete the work at the terminal. The terminal is the window that entered the command, ran the program.
In the official summary, Kimi K3.1 Pro was 65.7 per cent, Claude Opus 5.5 66.4 per cent and GPT-6 Astra 58.2 per cent.
In this case, Kimi and Opus were only 0.7 percentage points away.
By 3D modeling, it changes.
BenchCAD will give AI a few different angles to write CAD codes and make models.
Kimi achieved 92.7 per cent, Opus 96.2 per cent and Astra 95.9 per cent when it was allowed to call. The with alls in the graph, is that we can use tools.
Kimi also made a set of internal comparisons, focusing on the long tasks of multiple steps.
The test, Kimi Code Bench 3.0 Long-horizon, focused on the ability of AI to do many steps in a row, to use tools, and to complete the work.
K3 is 55.95% and K3.1 Pro has reached 61.2%, an increase of 5.25 percentage points.
K3.1 Pro continues to be trained on the basis of K3, with 1 million Tokencontext in previous generations.
Finance model, formula and projections delivered
In addition to the games and animations, several official documents were submitted: the financial valuation model of Ingweida, a presentation by the Nindian Commission and a report from the quantum computing industry.
The committee, the group that discussed whether to invest, met. The presentation needs to clarify the judgement and basis.
The key is in the table.
A few tables are connected, and the formulas can be viewed. You change one of the numbers, the results change.
This balance sheet below shows the actual data, projections column and formulae.
The study goes on to take on another matter: to see a judgement, and to quote it, to find the original information it uses.
The tables can be continued and the report can be reviewed. Those who make materials every day should know how much it saves.
On OfficeQA Pro v2, K3.1 Pro's official achievement was 64.4. This test will allow AI to look for information from authentic financial documents and then complete the calculations against several documents.
At the end.
In the past, there was a world in one's head, and in the end there was probably only one text left. He doesn't have to be the one who wants to write stories, but to model, write programs, do animations, and learn from the beginning.
If these jobs are given more to AI, the order of choice may be reversed.
First of all, should the world be read, seen or walked in?
Thank you for reading my article.
If you think it's good, you can order it, read it, forward it.
If you want to get a delivery first, you can give me a sign. Thank you for reading my article.
Open-source knowledge base (real-time update exchange group):
my.feishu.cn/wiki/OhQ8wq…
Auto-fetched and translated by AI Atlas. All rights belong to the original author.