刚刚,Kimi 正式发布了新一代旗舰模型 K3.1 Pro。在国产大模型纷纷卷起速度和售价的当下,Kimi 又把竞争带回了表现天花板之争。

前两天,大模型竞技场上一款叫 Kibitz 的神秘模型,就已经引起了人们的关注。社交媒体的信息流里,都是用这款模型生成的各种惊艳的案例。

这个骑自行车的鹈鹕,一行提示词已经进化到在不同的环境背景里面骑行,还能调节步频。不少网友认为,成效已经足以媲美 GPT-6 Astra 和 Opus 等前沿闭源模型。
现在 Kibitz 的真实身份也揭晓了,它就是 Kimi K3.1 Pro。一起公开的,还有这款新模型在竞技场上的盲测成绩。

在代码、Agent 和知识类基准上,K31. Pro 都紧追最强闭源模型,在 Artificial Analysis 「智能指数」直接跃升到全球第四,开源第一。

除了保持 benchmark 的领先,K3.1 Pro 一如既往地没有挑选走两三个月来国模的 Flash 路线,它继续用旗舰大模型来占据国产模型 Pro 档位的第一。
K3.1 Pro 具备 3 万亿级参数模型,兼容 100 万上下文,并计划 11 月 4 日前开放完整模型权重。这次 Kimi 还同步推出了 K3.1 Pro Highspeed 高速版本,在保持相同的智能水平的前提下,输出速度能做到普通版的 2-3 倍。

笔者连夜测试了 K3.1 Pro 在代码、3D 创作、金融研究与建模等维度的能力,只能说那个「大而美」的 Kimi 时隔100天又回来了。
各种 3D 大场面,全部都能做
这段时间最「大」的场面莫过于「特斯拉能飞」了,笔者也是安排 K3.1 Pro 复刻一下特斯拉的空中特技。
模型利用 Three.js 和 SVG 制作了一段汽车短片,白色的 Roadster 冲上巨大的 C 形轨道,在最高处倒挂一秒,随后脱离轨道,喷出冷气,在半空翻回正面,最后四轮落地。

回到地面, 再让 K3.1 Pro 用一句话就造出一个 3D版「一剑冰封整座城」。
生成的雨夜古城里,玩家蓄力拔剑后,寒潮沿街巷向远处推进,把城楼、屋顶、树木和河面依次冻成透亮的冰晶。

K3.1 Pro 真正做到了「一句话就生成可玩特效」,情形的视觉冲击力、材质、灯光、空间层次和动作节奏做到了像正式游戏的大招展示。
古城之后,画面再转进墨西哥高地的蝴蝶森林,以帝王蝶的越冬栖息地为主题,生成一个虚拟的 3D 情形。

蝴蝶给画面带来了几种不同的尺度,远处是压满枝头的色块,靠近以后是层叠的翅膀,再近一些,才看清单只蝴蝶的纹样。透过翅膀的光,也让这个 3D 案例有了和古城不同的质感。
除了让 K3.1 Pro 生成这些 3D 大场面,Kimi 这次还同步上线了一个挺有意思的新能力:3D Make。笔者可以直接在线预览、编辑 3D 模型,调整好之后,模型文件还能下载下来,拿去 3D 打印。

如果不打算打印,3D Make 也兼容把作品导出为视频或实况照片,直接分享到社交平台,可太合适 AI 玩家了。
从审美到工程,Kimi 的另一面
从 3D 项目成为今年大模型的必测项目以来,能看到模型在 3D 成效上的表现都越来越好,更多的细节表现,渲染也更高质,能体现不同模型区别的,更多的还是在模型的审美上,或者说模型的「品味」。
网页,这项曾经是 Kimi 最具长处的项目,随着 K3.1 Pro 的能力提高,现在笔者用更简单的方式,更快得到一个真正有外观感的网页。

类似的能力,给 K3.1 Pro 一张房间照片,它能把照片里的空间变成一个可以旋转浏览的 3D 网页。
桌子、椅子、书架和摆件先从原位分离,桌板和桌腿、椅背和座面、显示屏和支架再逐步展开。最后,所有物件像拆机摄影那样,整齐铺成一个矩形。

又或者让 K3.1 Pro 用 Blender 脚本生成一个 3D 机器人资产,然后再为这款掌心大小的桌面陪伴机器人做一场可交互的 3D 发布会。

K3.1 Pro 生成的网页,把材质、镜头和排版都做到了尽可能精致,同一套模型在这里既能演示,也能继续交给外观师修改。
最近用大模型做出来比较好玩的还有解释机器,笔者丢给它一个「XX 是怎么工作的」难题,它就把答案做成一台可以点的互动机器,用一个独立的 HTML 页面,把真实的部件、可以按步骤点亮的按钮、以及能剖开看内部运作的窗口都囊括进来。
发送提示词「冰箱是怎么制冷的」让 K3.1 Pro 做一个网页,生成的成效很合适用来做简单的科普。

还有前段时间有个做插画的 Skill 很火,当时的教程建议使用 Claude Code 的模型和工具。笔者也把它安装到 Kimi 里,看看它能不能也生成像 Opus 5.5 一样高水准的动效插画网页。

Anatomy 这个 Skill 的特征也是把工艺概念做成一台「真的机器」来解释,它几乎不会使用画方框和箭头的方式,而是发明一个物理装置,让它的工作原理本身就是那个概念。
然后笔者在测试 K3.1 Pro 的时候,给它上了点难度,要求它来解释一下海底光缆是怎么连接全球互联网。

事实证明,在处理这类有专门 Skill 的创作任务时,K3.1 Pro 完全可以用来替代 GPT-6 Astra 或 Opus 5.5 等前沿旗舰模型,笔者可以花更少的钱就得到同样好的成效。
除了 3D 能力的提高,笔者还发现,K3.1 Pro 经过大规模强化学习之后,长程编码能力也有了显著增强。
例如在测试中,笔者顺手让它做了一段动态图形视频。提示词只有一句话:「制作一段充满动感的 15 秒动态图形视频,展现你是一位多么出色的动态外观师,就像这是你简历用的作品集短片一样。」

最终生成的成品,已经有了比较完整的节奏、转场和视觉语言。
Kimi 在通用编程和长程调试上的增强,不只体现在创意生成任务里,也体现在更复杂的软件理解和复现任务上。

刚好 Arc 将在 12 月关闭,笔者就想着安装好 REA 后,让 K3.1 Pro 复刻一个同款的 Arc 浏览器。
能看到 K3.1 Pro 对安装包开展了完整的检查,提取了详细的应用信息。而复刻的版本,笔者也能像用 Arc 一样,使用标签页拆分、侧边栏、全局搜索以及空间等能力。

除了好看,还很专业
3D、网页和动效这些案例,最容易让笔者看到模型的进步。但对于真正需要用 AI 干活的企业来说,模型能否正确处理复杂数据、完成专业任务,可能比做出一个漂亮的网页更关键。
金融就是一个很好的例子,如果让 AI 总结一份财报很简单,但要同时读取多份金融文件,交叉核对数据,再建立一份包含大量公式的财务模型,难度就高了不少。
依据 Kimi 公布的成绩,这次 K3.1 Pro 在金融领域的表现也有显著提高。在 Office QA Pro v2 和 SpreadsheetBench 2 两项外部基准测试中,K3.1 Pro 均超过了 GPT-6 Astra、Opus 5.5 等前沿闭源模型。

前者考察模型对真实金融文档的分析能力,后者则要求 AI 真正操作电子表格,涵盖建立财务模型、处理跨表公式和诊断错误。
国产模型终于继续卷 Pro 了
下半年大模型更新基本收缩成两条主线,卷智价比的「斩杀线」,以及顶尖的前沿能力。
从 DeepSeek V4 Flash 开始,人们不约而同地把重点放在了卷成本和速度的 Flash 上。
另一边,Claude Opus 5.5、GPT-6 Astra 等前沿模型接连更新,尤其是在编程、3D 创作和复杂 Agent 任务上,表现一次比一次惊艳。
Kimi K3 在今年 7 月发布时,把国产开源模型的规模提高到了 2.8 万亿参数。当时它最受关注的一点,是让开源模型的总体能力进一步逼近了美国顶级闭源模型。
但模型的迭代速度实在太快,短短两个月,很多 Flash 模型已经达到甚至超越之前的旗舰模型,「斩杀线」不断刷新,而 Claude 则稳稳占据的前沿模型的头把交椅。

现在,Kimi 终于带着 K3.1 Pro 回来了,它当然还没有全面超过这些闭源旗舰,但从这次的实测来看,至少国产模型这一边,终于又有选手能和它们正面较量了。
某种程度上,Kimi K3.1 Pro 也代表着一条「大而美」的工艺路线。
过去一段时间,大模型行业热衷于证明一件事,小模型也能办大事。借助工艺工程,更少的参数、更低的推理成本,也能在 Benchmark 上交出接近旗舰模型的成绩单。
但 Benchmark 上分数接近,不代表现实世界里的能力就完全相同。
类似的差异在开放式任务中尤其显著。两款模型即便跑分相差无几,真正拿来写文章、做创意策划或者分析复杂难题时,较大模型仍然可能展现出更敏锐的洞察力、更多元的想象力,以及更细腻的语言表达能力。

参数规模并不能直接等同于智能水平,但在架构、训练质量和推理条件相近的情况下,更充足的模型容量,会带来知识储备和综合能力上的长处。
面对大型代码工程、复杂 Agent 工作流等现实任务,模型还需要持续规划、调用工具、应对突发状况,并在漫长的执行过程中始终理解自己究竟要完成什么。相比标准化测试,真实世界对模型的要求显然更加复杂。
说到底,模型可以越来越便宜,简单任务也可以交给越来越小的模型,但真到了需要干大活儿的时候,使用者仍旧需要一个能力上限足够高的 Pro。
从这个角度来看,Kimi K3.1 Pro 的意义也正在于此。当行业都忙着把模型做得更快、更省时,仍然有人愿意继续探索大模型的能力上限,让国产模型有机会参与全球最顶尖的能力竞争。
说实话,能重新对一款国产旗舰模型有这样的期待,还是挺让人兴奋的。
本文来源:https://www.ifanr.com/1685684
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Just now, Kimi officially released the next-generation flagship model, the K3.1 Pro. Now that the big domestic models are rolling up the speed and price, Kimi has brought the competition back to the performance ceiling.
Two days ago, a mysterious model called Kibitz in the big model arena had already attracted people's attention. Social media feeds are full of stunning examples generated with this model.
This bicycle riding pelican, a line of prompts has evolved to ride in different environmental backgrounds, and can also adjust the pace. Many netizens believe that the results are enough to rival the cutting-edge closed-source models such as GPT-6 Astra and Opus.
Now that the true identity of Kibitz has been revealed, it is the Kimi K3.1 Pro. Together with the public, there are blind tests of this new model on the arena.
On code, agent, and knowledge benchmarks, K31. Pro is closely following the strongest closed-source model, and has directly jumped to the fourth in the world in the Artificial Analysis "Intelligence Index" and the first in open source.
In addition to maintaining the lead of benchmark, K3.1 Pro has not selected the Flash route of the national model for two or three months, and it continues to occupy the first position of the domestic model Pro with the flagship large model.
K3.1 Pro has a 3 trillion-level parameter model, compatible with 1 million contexts, and plans to open full model weights by November 4. This time, Kimi also launched a high-speed version of the K3.1 Pro Highspeed, which can output 2-3 times faster than the normal version while maintaining the same level of intelligence.
I tested the ability of K3.1 Pro in the dimensions of code, 3D creation, financial research and modeling overnight. I can only say that the "big and beautiful" Kimi came back after 100 days.
The biggest "big" scene during this period was "Tesla can fly". I also arranged for the K3.1 Pro to reproduce Tesla's aerial stunts.
The model uses Three.js and SVG to make a short car film. The white Roadster rushes on the huge C-shaped track, hangs upside down at the top for one second, then disengages from the track, ejects cold air, flips back to the front in midair, and finally lands on the ground with four wheels.
Return to the ground and let K3.1 Pro create a 3D version of "one sword frozen the whole city" in one sentence.
In the rainy night generated in the ancient city, after the player plucked the sword, the cold tide pushed along the streets and alleys to the distance, freezing the city towers, roofs, trees and river surfaces into transparent ice crystals in turn.
K3.1 Pro really does "generate playable special effects in one sentence", and the visual impact of the situation, the material, the lighting, the spatial level and the action rhythm are displayed like the big moves of the official game.
After the ancient city, the picture is turned into the butterfly forest in the highlands of Mexico, and a virtual 3D situation is generated with the theme of the winter habitat of the monarch butterfly.
The butterfly brought several different scales to the picture. In the distance, there were color blocks full of branches. Closer, there were layers of wings. Closer, I could only see the pattern of the butterfly in the list. Through the light of the wings, this 3D case also has a different texture from the ancient city.
In addition to letting K3.1 Pro generate these 3D big scenes, Kimi also launched a very interesting new capability this time: 3D Make. The author can directly preview and edit the 3D model online. After adjusting, the model file can be downloaded and 3D printed.
If you don't plan to print, 3D Make is also compatible with exporting your work as a video or live photo and sharing it directly to social platforms, which is great for AI players.
Since the 3D project has become a must-test project for this year's large models, we can see that the performance of the models in 3D performance is getting better and better, with more detailed performance and higher quality rendering, which can reflect the differences between different models, and more still in the aesthetics of the models, or the "taste" of the models.
Webpage, once Kimi's strongest project, with the improvement of K3.1 Pro capabilities, now I use a simpler way, faster to get a really look and feel of the webpage.
Similarly, give the K3.1 Pro a room photo, which turns the space in the photo into a 3D webpage that can be rotated and browsed.
Tables, chairs, bookshelves and ornaments are separated from the original position, and the table plate and table legs, chair back and seat surface, display screen and bracket are gradually unfolded. Finally, all the objects are neatly laid out in a rectangle like a disassembled camera.
Or let K3.1 Pro generate a 3D robot asset with Blender scripts, and then do an interactive 3D launch for this palm-sized desktop companion robot.
The webpage generated by K3.1 Pro makes the material, lens and typography as refined as possible. The same set of models can be demonstrated here and can continue to be modified by the designer.
Recently, it is more fun to use a large model to make an explanation machine. The author threw it a "how does XX work" puzzle. It makes the answer into an interactive machine that can be dotted, and uses a separate HTML page to include the real parts, buttons that can be lit step by step, and windows that can be opened to see the inner workings.
Send the prompt "How is the refrigerator cooled?" Let K3.1 Pro make a webpage, and the generated results are very suitable for simple popularization of science.
There was also an illustrator, Skill, who was hot some time ago, and the tutorial at that time recommended using Claude Code's models and tools. I also installed it in Kimi to see if it can also generate animated illustration web pages as high as Opus 5.5.
The Skill of Anatomy is also characterized by an explanation of the concept of craftsmanship as a "real machine", which hardly uses boxes and arrows, but invents a physical device, so that its working principle is that concept itself.
Then, when I tested the K3.1 Pro, I gave it a little difficulty and asked it to explain how the submarine fiber optic cable connects to the global Internet.
It turns out that K3.1 Pro can be used to replace cutting-edge flagship models such as GPT-6 Astra or Opus 5.5 when dealing with such creative tasks with specialized skills, and I can get the same good results for less money.
In addition to the improvement of 3D capabilities, the author also found that the long-range coding capabilities of K3.1 Pro have also been significantly enhanced after large-scale reinforcement learning.
For example, in the test, the author handily let it make a moving graphic video. The prompt has only one sentence: “Create a dynamic 15-second motion graphic video that shows what a great dynamic visualizer you are, like this is a portfolio clip from your resume.”
The resulting finished product has a relatively complete rhythm, transition and visual language.
Kimi's enhancements in general-purpose programming and long-range debugging are reflected not only in creative generation tasks, but also in more complex software understanding and reproduction tasks.
Just in time for the closure of Arc in December, I thought that after installing REA, I would have the K3.1 Pro reproduce an Arc browser of the same model.
You can see that K3.1 Pro has carried out a complete inspection of the installation package and extracted detailed application information. In the replica version, I can also use the ability to split tabs, sidebars, global search, and space like with Arc.
3D, web and animation cases are the easiest for the author to see the progress of the model. But for enterprises that really need to work with AI, whether the model can correctly handle complex data and complete professional tasks may be more critical than making a beautiful web page.
Finance is a good example. If you let AI summarize a financial report, it is very simple, but it is more difficult to read multiple financial documents at the same time, cross-check the data, and build a financial model with a large number of formulas.
According to the results released by Kimi, the performance of K3.1 Pro in the financial field has also improved significantly. In the two external benchmark tests of Office QA Pro v2 and SpreadsheetBench 2, K3.1 Pro surpassed the leading-edge closed-source models such as GPT-6 Astra and Opus 5.5.
The former examines the model's ability to analyze real financial documents, while the latter requires AI to actually operate spreadsheets, covering building financial models, handling cross-table formulas, and diagnosing errors.
In the second half of the year, the large model update basically shrunk into two main lines, the "kill line" of the intelligence-price ratio, and the cutting-edge capabilities.
Starting with DeepSeek V4 Flash, people have been focusing on Flash for volume cost and speed.
On the other hand, cutting-edge models such as Claude Opus 5.5 and GPT-6 Astra have been updated one after another, especially in programming, 3D creation and complex Agent tasks.
When Kimi K3 was released in July this year, the scale of the domestic open source model was increased to 2.8 trillion parameters. At that time, it was most concerned about the overall ability of the open source model to further approximate the top closed source model in the United States.
However, the iteration speed of the model is too fast. In just two months, many Flash models have reached or even surpassed the previous flagship model. The "kill line" is constantly refreshed, and Claude is the leading edge model that steadily occupies the top spot.
Now, Kimi has finally returned with the K3.1 Pro. Of course, it has not completely surpassed these closed-source flagships, but from the actual test this time, at least on the domestic model side, there are finally players who can compete with them positively.
In a way, the Kimi K3.1 Pro also represents a “big and beautiful” process.
In the past, the big model industry was keen to prove that small models can do big things. With the help of process engineering, fewer parameters, lower inference costs, and transcripts close to the flagship model can also be handed over on Benchmark.
But a close score on the Benchmark does not mean that the abilities in the real world are exactly the same.
Similar differences are particularly noticeable in open-ended tasks. Even if the two models are almost the same, when they are actually used to write articles, do creative planning, or analyze complex problems, the larger model may still show more keen insight, more diverse imagination, and more delicate language expression.
Parameter size is not directly equivalent to intelligence level, but when the architecture, training quality and inference conditions are similar, more sufficient model capacity will bring the advantages of knowledge reserve and comprehensive ability.
In the face of real-world tasks such as large-scale code engineering and complex Agent workflows, the model also needs to continuously plan, call tools, deal with unexpected situations, and always understand what it is about to accomplish during the long execution process. Compared to standardized testing, real-world requirements for models are clearly more complex.
In the end, models can get cheaper and cheaper, and simple tasks can be handed over to smaller and smaller models, but when it comes time to do big work, users still need a Pro with a high enough capacity limit.
From this perspective, that's what the Kimi K3.1 Pro is all about. When the industry is busy making models faster and less time-consuming, there are still people who are willing to continue to explore the upper limit of the ability of large models, so that domestic models have the opportunity to participate in the world's top competitiveness.
To be honest, it is quite exciting to have such an expectation for a domestic flagship model again.
This article is from: https://www.ifanr.com/1685684 For copyright issues, please write to: 13951431913#139.com or submit a review on the right.