新智元报道
昨晚的X，被刚刚发布的Kimi K3.1 Pro正式刷屏了！
外国开发者们惊呼：欢迎来到3D Web 开发的新时代。
之前，AI榜单前十几乎被美国公司包圆了。相对而言，国模这边有点沉寂。
有人感慨：国产大模型最近发的都是主打性价比的「Flash」模型，真能打的「Pro」去哪了？
很多人在问：国模是不是掉队了？
今天，Kimi打破了沉默——Kimi K3.1 Pro来了！
知名AI评测机构Artificial Analysis也公布了成绩：K3.1 Pro以55分冲上全球智能指数第4，较前代暴涨11分。
在考验Agent实战能力的Terminal-Bench 4.0上，它更是从13%飙升至65.7%， 距离 Opus 5.5 不到 1%！
Artificial Analysis还给出一个重磅判断：一旦权重发布，K3.1 Pro将登顶开放权重模型榜首，开源与顶级闭源模型的差距，将从12分骤缩至3分！
今天早上，reddit已经被这个跑分刷屏：Kimi K3.1超越了Fable 5.1！
国模这次，真把全球第一梯队的门踹开了。
这次Kimi端上来的，不是个过渡版本，而是一盘硬菜。
这是迄今为止能力最强的 3 万亿级参数模型，支持高达 100 万上下文。
它的出现告诉世界：China is back，国模王者归来！
在各大硬核测试中，K3.1 Pro 的性能直接比肩海外顶流 GPT-6 Astra！更离谱的是，它的成本只有 Astra 的五分之一，便宜了整整 80%！
所以，K3.1 Pro 的定位很清楚：做前沿模型里，最理性的那个选择。
而且，在真实场景中，要干有挑战的任务，还是得靠更大的模型。举个例子，尽管基准测试分数跟Opus相近，但真到了现实场景里，Fable 这种较大模型，经常会表现出更强的洞察力、创作力和写作能力。
毕竟，高难度推理任务，才能显出大模型的本事。
这一次，K3.1 Pro 基于 K3 做后训练，月之暗面在强化学习上大幅加码算力。
为了让几百个下游领域一起联合训练，他们从 RL 框架、长上下文推理、沙箱，到调度系统、环境质检平台，整套技术栈逐层重新打磨。目标就是——每一份新增的算力，都要变成实打实的能力。
结果是，在代码、Agent 和知识类基准上，K3.1 Pro 达到或接近最强闭源模型的水平。
在考察 Agent 在真实终端里独立干活的 Terminal-Bench 4.0 上，它拿到了65.7分，而且在三种思考强度下，都落在「性能与单任务成本」的帕累托前沿上。
也就是说，同样的钱，它干得最多；同样的活，它花得最少。
Kimi 还贴心地端出了高速版模型 Kimi K3.1 Pro Highspeed，智商不减，但输出速度直接飙升 3 倍。
不过，跑分只是门票。榜上分数咬得再紧，一到真刀真枪的活儿上就会拉开差距——
任务越难，越是拼模型本身的底子，这恰恰是大模型的主场。
接下来，我们就看看 K3.1 Pro 在真实工作流里，是怎样表现的。
据悉，K3.1 Pro 有个引人注目的变化：审美，成了模型的能力项。
Kimi想回答的问题是——代码模型，能不能好看？
这一次，无论是前端页面、Three.js 交互场景，还是3D 资产，它在构图、排版、动效和完成度上都有明显提升。
它还会用 Unity、Unreal Engine、Godot 这些游戏引擎，可以不靠 MCP 连接器，自己写脚本直接驱动 Blender、SolidWorks、Rhino 建模。
接下来，就让我们来一波硬核实测，来看看「大而美」的K3.1 Pro，究竟有多能打？
这是K3.1 Pro主打的新技能。官方说它能从一段文字、一张草图或参考图推断形状、比例和空间关系，直接写 Blender 脚本建模，再接入 Three.js 或游戏引擎。
我们来验证一下，这条链路是不是真的能打通。
一张手绘三视图扔进去，40 分钟，它交出来的是：608 行 Blender 脚本、可编辑的 .blend、1.7 MB 的 GLB，还有一个双击就能开的网页。
草图上写的翼展 8.0、总长 6.2、总高 2.4，建出来是 8.000、6.199、2.400，差了 1 毫米。
接下来，它开始了丝滑的操作。
动手前，K3.1 Pro会先试探 Blender 5.2 的新接口，确认了再写正式脚本；验收标准直接写成断言，旋翼直径差 1 毫米就不许导出；渲染完自己对照草图改了一轮；网页做完，它甚至还开浏览器给自己截了两张截图，这样才算交差。
六项待办，自己一项项划掉。
机身是样条插出来的流线体，座舱玻璃是吸附在机身上的曲面，桨叶带扭转，四个旋翼的倾转层级从 Blender 一路到浏览器里，滑块拉到底，真的绕支臂往前倒 90°。
全程，K3.1 Pro都没碰 MCP，没用外部模型库。
写到一半时，1M 上下文才用了 6%。40 分钟不是它的上限，是这张草图只有这么难。
更令人振奋的是——这套活儿不用本地装 Blender 也能干！
kimi.com 和 Kimi app 已同步上线 3D Make 功能。把同一份草图和资料直接丢进 Kimi，一句「参考这些资料，把它的 3D 模型做出来」，它就在云端开工了，这次又跑了将近 40 分钟。
改 scene.js、改 style.css，跑终端重建，它发现两片桨叶转起来不对称，改成半整数转速；CSS 动画暂停不一致，修；截图时场景还没加载完，加一个「就绪」再截。
最后自己开浏览器点了一遍，报了句「无脚本错误或资源缺失」，才把版本号交出来。
做完的东西就在右边窗口里，能转、能换色、能拉倾转滑块，不满意直接在对话里改，它接着改。
使用3D Make可以在线预览和编辑，一键发布成网页，GLB 下载下来可以直接去 3D 打印，还能导出视频或实况照片发社交平台。
https://lsx4v6wl2amc4.kimi.page/
以前，这是建模师、前端和运营三个人的活，现在，K3.1 Pro用一个输入框就完成了。
而众多网友的实测也证明，K3.1 Pro的审美绝了。
仅用4.92美元，它就在一款绘画应用程序中画出一副肖像。
有人在虚幻引擎中建造了一座云端宫殿，成本不到10美元，引发了众多网友盛赞。
鹈鹕骑自行车的demo，审美绝了。
以前的AI碰金融，只会写摘要，一遇到跨表公式和估值模型就胡言乱语。
但 K3.1 Pro 这次，直接干翻了海外顶流。
在 OfficeQA Pro v2（基于真实金融文档的复杂财务数据分析，考察检索、跨文档核对与多步数值推理）上，K3.1 Pro 拿下 64.4 分，当前最优（SoTA）。
在覆盖从标的筛选到盈利分析整条投资工作流的 Frontier Finance 上，K3.1 Pro 位列第二，较 K3 进一步提升。
接下来，我们看看 K3.1 Pro 在真实投行工作流里，是怎样表现的。
它给我们手搓了一份英伟达的专业级财务模型。
它并不是简单地填表，而是能端到端交付一整个带公式联动的 Excel！它像极了老练的分析师，经营驱动因子可随情景调整，敏感性分析直接呈现对估值的影响，每一项假设都有硬核数据支撑。
可以说，现在的K3.1 Pro是一个真正懂财务逻辑的「全栈金融民工」了。
Kimi K3.1 Pro 的出现，证明了我们在大模型的牌桌上，依然握有随时打出王炸的实力。
它当然不是全能冠军，在一些维度上离最强闭源模型还有距离。
但真正令人振奋的，是国模正在向全球智能的性能天花板发起正面冲锋。
性价比决定市场，能力上限决定未来。
Kimi K3.1 Pro已经能与全球最强模型正面交锋，甚至重新定义智能的边界。
而且这次性能更强了，价格反而更便宜：比 K3 便宜了三分之一，只有 GPT-6 Astra 的五分之一。
现在，不用苦等解封了。今日起，把Kimi K3.1 Pro用起来。
China is back！
编辑：所罗门 Aeneas
不喜欢
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。
New Ji-won.
Last night's X was officially painted by Kimi K3.1 Pro of publish!
Foreign developers are crying: Welcome to the new era of 3D Web development.
Before, the top 10 on the AI list was almost rounded up by American companies. In relative terms, it's a bit quiet on this side of the country model.
Some are outraged that LLM, the national product, recently issued a "Flash" model, which is price-driven.
A lot of people are asking, "Is the national model out of line?"
Today, Kimi broke the silence - Kimi K3.1 Pro came!
Artificial Analysis, a well-known AI rating agency, also published results: K3.1 Pro hit Global Intelligence Index 4 with 55 points, 11 points higher than previous generations.
At Terminal-Bench 4.0, which tested Agent's operational capability, it jumped from 13% to 65.7%, less than 1% from Opus 5.5!
Artificial Analysis also makes a heavy judgement: once the weight is publish, K3.1 Pro will top the top of the open weight model, the gap between the open source and the top closed source model will shrunk from 12 to 3!
This morning, reddit has been drawn from this run: Kimi K3.1 surpassed Fable 5.1!
National Moot really kicked the door open for the first team in the world.
This time Kimi brought up not a transitional version, but a hard plate.
This is the most capable 3 trillion-degree parameter model to date, supporting up to 1 million contexts.
His presence tells the world that China is back, the King of the Kingdom!
In all major hard core tests, K3.1 Pro's performance is directly compared to the top of the shoulder overseas GPT-6 Astra! And even worse, it cost only one fifth of Astra's price, and it's down 80%!
So K3.1 Pro is clearly positioned: the most rational option in the front-line model.
Moreover, in the real scene, more challenging tasks still depend on larger models. For example, even though the benchmark scores are close to Opus, in reality, Fable, like LLM, often shows greater insight, creativity and writing skills.
After all, the difficult task of reasoning is to show the ability of LLM.
This time, K3.1 Pro is based on K3 doing post-training, and the dark side of the moon adds a lot of weight to enhanced learning.
In order to bring together hundreds of downstream areas for joint training, they have gone from the RL framework, long context reasoning, sandboxes, to the movement control system, the environmental quality control platform, and the entire technology stacks are being reworked layer by layer. The goal is — every additional calculus must be translated into real capacity.
As a result, K3.1 Pro reached or was close to the level of the most powerful closed source model in code, Agent and knowledge type benchmarks.
Agent received 65.7 points on the Terminal-Bench 4.0, which worked independently at the real terminal, and fell on the front line of the Paretto "performance and single job cost" under three types of thinking.
That is, the same money, it does the most; the same thing, it costs the least.
Kimi also carried out the high-speed model Kimi K3.1 Pro Highspeed with a steady IQ, but the output went up three times faster.
However, running points are just tickets. The score on the list is even tighter, and once you get to the job of real guns...
The harder the task, the more it is to spell the foundations of the model itself, which is precisely the main site of LLM.
Next, let's look at K3.1 Pro in real workflow.
K3.1 Pro is known to have a remarkable change: aesthetics, becoming a model's ability item.
The question Kimi wanted to answer was, "Code model. Can it look good?"
This time, whether the front-end page, the Three.js interactive scene or the 3D asset, there was a marked increase in the image, layout, movement and completion.
It also uses the Uniity, Unreal Engineering, Godot engines, which can drive the Blender, SolidWorks, Rhino without MCP connection.
So, let's take a hard check to see how big and beautiful the K3.1 Pro is.
This is K3.1 Pro's new skills. Officially, it can extrapolate shapes, scales and spatial relationships from a text, a sketch or a reference map, directly write a Blender script model, and then access Three.js or a game engine.
Let's check that this link is really working.
A hand-drawing three view was thrown into it for 40 minutes, and it gave out: line 608, Blender script, editable.blend, 1.7 MB GLB, and a double-clickable web page.
Wing show 8.0, General Commander 6.2, Total height 2.4, constructed 8.000, 6.199, 2.400, 1 mm.
Next, it starts the silk slide.
Before doing so, K3.1 Pro will test the new interface of Blender 5.2 and confirm and write the official script; acceptance criteria will write the assertion that the rotor's diameter is less than 1 mm and will not be exported; refill itself with a round of contrasting sketches; and, after the completion of the web page, it even opens its browser to intercept two screenshots of itself, which is the difference.
Six to-dos, one item per item.
The fuselage is a type of fluid that is plugged in, the cabin glass is the curve attached to the machine, the oars are reversed, and the tip of the four rotary wings goes all the way from Blender to the browser, and the slide is pulled to the bottom, and it really turns around the arm 90 degrees.
All the way, K3.1 Pro didn't touch MCP, no external model library.
In the middle of it, it was only 6% of the 1M context. Forty minutes is not its limit, but this sketch is so difficult.
And what's even more exciting is that this job can be done without a local costume.
Kimi.com and Kimi app are on line with 3D Make. Throwing the same sketches and data directly into Kimi, and saying, "Refer to this information, make its 3D model," it started at the cloud end, and this time it ran for almost 40 minutes.
Changed scene.js, changed syle.css, rebuilt running terminals, and found that two oars were asymmetrical and converted to semi-integer speed; the CSS animations were suspended and repaired; the scene was not loaded and a "ready" was added.
Finally, the browser was set up and the version number was handed over with the phrase "no script error or lack of resources".
What is done is in the right window, which turns, changes colour, pulls slides, is unsatisfactory and changes directly in the conversation, and then it changes.
3D Make allows online previewing and editing, with publish as a web page, GLB downloads directly to 3D printing, and exports video or live photos to social platforms.
https://lsx4v6wl2amc4.kimi.page/
Previously, this was the work of a modeler, front end and operator, and now K3.1 Pro has completed it with an input box.
And so many of our online users have confirmed that K3.1 Pro's aesthetics are perfect.
For $4.92 only, it draws a portrait in a painting application.
The construction of a cloud palace in an illusory engine, at a cost of less than $10, has aroused a great deal of popularity.
Peking demo, aesthetic.
The former AI was only able to write summary, and once it encountered cross-table formulas and valuation models, it was nonsense.
But K3.1 Pro, this time, went straight to the top of overseas.
On OfficeQA Pro v2 (complex financial data analysis based on genuine financial documents, search, cross-document reconciliation and multi-step numerical reasoning), K3.1 Pro takes 64.4 points, currently best (SoTA).
K3.1 Pro ranks second, further up K3 on Frontier Finance, which covers the entire investment workflow, from the filter of the object to the profit analysis.
Next, let's see how K3.1 Pro acts in the real investment bank workflow.
It gave us a professional financial model for Weidar.
It's not simply a form, it's a formula that can be delivered to the end, Excel! It resembles a sophisticated analyst, with a business driver that can be adjusted to the situation, sensitivity analysis that directly presents the impact on valuation and each assumption is supported by hard core data.
It can be said that K3.1 Pro is now a financial domestic worker who really understands financial logic.
The appearance of Kimi K3.1 Pro is proof that we still have the power to hit Wang at LLM table.
It's certainly not an all-powerful champion, and in some dimensions it's far from the strongest closed source model.
But what is really exciting is that the national model is launching a positive push towards the global smart performance ceiling.
The value for money determines the market and the capacity ceiling determines the future.
Kimi K3.1 Pro has been able to confront the world's strongest models and even redefine intelligent boundaries.
And this time it's better, it's cheaper: one third cheaper than K3, one fifth cheaper than GPT-6 Astra.
Now, don't wait to untie it. From today, use Kimi K3.1 Pro.
China is back！
Editor: Solomon Aeneas
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.