15 分钟交付工业级可交互3D 资产，重写视觉大模型的“理科生法则”
作者丨高允毅
编辑丨岑 峰
刚刚，Kimi 官方发布了一段“极具暗示性”的先导片，各种高能科学画面与粒子特效快速切换，有眼尖的网友已经捕捉到画面中的彩蛋“Kimi K3.1”。
Kimi K3.1 要来了？这次它带着极强的视觉效果来了，有网友指出官方视频就是 Kimi K3.1 做的。
图注：眼尖的网友在画面中找到了亮点“Kimi K3.1”AI 博主 Chris 不仅肯定了即将发布的消息，而且表示他已经实测了一周。他透露，Kimi K3.1在开源 3D 图形库 Three.js 上表现出色，而且这个模型的成本只有 Kimi K3 的 70%，Token 效率全面升级。从 Chris 晒出的实测画面中，可以看到 Kimi K3.1 跑完了一整套 24 小时昼夜环境模拟，从晨光漫射到夜幕沉落，光影过渡帧级精准，环境细节随光线同步流转，全程丝滑无断点。要知道，Three.js 编写的程序是非常消耗资源的，因为它需要实时计算 3D 空间中的顶点、光影和纹理。而 Kimi 这波相当于用纯代码跑出了渲染级画质。实际上，关于 Kimi K3.1 的消息早已有更多细节流出。9 月底，就有开发者在月之暗面 API 平台的配置文件里，发现了“kimi-k3-1”的模型标识。10 月 21 日，Kimi Code 的模型切换列表中，悄悄新增了 K3.1 选项，相当于官方提前给开发者开了小范围测试。10 月下旬，竞技场突然冒出两个战斗力极强的神秘模型，代号 Kibitz 和 Kalmia。其中，Kibitz 在前端与 3D 代码类任务中表现强势。在 "鹈鹕骑自行车" 这类考验空间想象力的渲染任务中，它凭出色的几何和物理直觉，生成了近乎完美的 3D 动态画面与矢量图；在其他的 3D 体素（Voxel）渲染盲测中，又靠对超长指令的精准控制力胜出。值得注意的是，在生成同等复杂度的 3D 代码时，它的 Token 消耗只有国外顶级模型的几十分之一。很快就有开发者通过分词器特征检验，推测出 Kibitz 就是 Kimi K3.1。这完全符合 Kimi 之前的命名规律，K2.5 叫 Kiwido，K2.6 叫 Kiwire，K3 叫 Kivine，K3.1 对应 Kibitz。随后 SGLang 代码库的一个合并请求，又意外泄露了更多细节：K3.1 新增了“Medium 中等思考”分级，思考强度由三档扩为四档，可根据任务难度灵活调度算力，兼顾速度和深度。在 Arena 盲测中，Kimi K3.1 把网页前端与图形渲染上的优势继承下来并进一步放大，在 3D 图形（Three.js/ WebGL）、交互大屏、H5 动态仿真等 6 大细分领域，展现出一定统治力。网上已经有不少一手内测案例流出，网友惊呼："这是一头 3D 猛兽”。要检验大模型的视觉与空间推理能力，AI 圈有个绕不过去的测试“鹈鹕骑自行车”。这个题目看似简单，实则处处是坑。因为这个画面无法参考现成的鹈鹕骑车素材，必须同时理解鹈鹕的解剖结构和自行车的机械逻辑，再把两个完全不相关的概念组合起来，用纯代码算出几何位置。这要考验模型的“概念组合能力”、“物理直觉能力”、“空间几何能力”。早期 Claude 3.5 画出的自行车只是两个圈加个三角，鹈鹕就是个带香蕉嘴的圆球，堪称 “抽象派代表作”。而 Kimi K3.1 的实测结果，直接把难度拉满。图注：早期 Claude 3.5 画的“鹈鹕骑自行车”在网友的实测案例中可以看到，Kimi K3.1 做了一个自行车骑鹈鹕的荒诞堆叠，却符合一定的物理规律，重心压得住、车身立得稳。除了鹈鹕，夜空挂着新月，远处山丘轮廓衬着橙红色落日余晖，左侧红白条纹的灯塔射出光束，背景小屋亮着暖光。细节上，海浪、小螃蟹、远处帆船剪影一应俱全。整个画面构图完整、光影层次丰富，所有物体的空间关系完全正确。更关键的是，这不是像素生成的图片，是模型用代码一行行算出来的矢量渲染，就像用数学公式完美还原了整个童话场景。对此，网友直言：“这种堆叠构图+光影层次+物体关系理解，已经不是简单SVG能随便糊弄的水平，完成度直接比肩 Opus 5.5，国产模型的前端视觉能力已经杀进第一梯队。” 有网友直接用 Kimi K3.1 生成了一座中世纪城堡，效果十分惊艳。城堡的细节设计非常丰富：主体是层层叠叠的多层塔楼，塔尖插着旗帜，在风中微微扬起；城门吊桥横跨护城河，周边环绕着群山、草地和错落的民宅。天空中还有飞鸟与浮云，光线被压成黄昏最后一抹暖色，打在石墙上形成明暗过渡，整座城堡仿佛 "在一天将尽时悬浮于空中"。更难得的是，悬浮的岩体、层叠的塔楼、依附的阶梯和植被，全都处在同一个空间逻辑里，彼此自洽。不过这位网友测完后表示“感觉有点像 GPT-6.1 Sol Max”。还有人指出，Kimi K3.1 新增了高级控制功能，有时间控制、渲染质量切换等等，能让场景效果很丰富。另一个案例是生成一整列能跑的蒸汽火车， Kimi K3.1不仅交付速度极快，而且完成度很高。尤其是，它还能把整列火车 "拆开" 给你看：车头、车厢、黑色的圆柱形储气罐、红色的车轮，所有部件都被拆解成独立零件，悬空错落地悬浮在画面里，每个零件形态完整、位置精确，像一张会动的“工程图”。从 3D 交互工程的角度看，Kimi K3.1 找到新的能力突破口。海外主流模型大多停留在静态建模和基础画面渲染，Kimi K3.1 则把结构设计、机械联动、交互逻辑连贯地串在一起，能完整的拆解视图、动态调速、实时数据追踪。这些说明模型读懂了物理工程原理，而不只是在堆视觉。它既保留了 GPT 式的通用理解，也具备 Anthropic 式的深度逻辑推演，这也许正是国产 AI 技术迭代的一条独特路径。有网友用 Kimi K3.1 ，15分钟就完成高精度、交互式3D直升机模型。懂行的人表示，这些都是宝贵的3D资产，能大幅压缩原型制作周期。鉴于 Kimi K3.1 出色的视觉表现，网友们忍不住拿它和全球顶尖视觉模型一一对比。在与 GPT-6 Astra 对比中，两个模型都在《我的世界》造建筑，一个搭出机器人造型的建筑，一个盖出球形建筑，似乎各有千秋，但 Kimi K3.1 的成本只有 GPT-6 Astra 的约 1/14。图注：GPT-6 Astra与 Kimi K3.1 挑战在《我的世界》造建筑在与 Haiku 5.5 的对比中，两个模型同时搭建村庄模型，都包含巨龙袭击、雷暴和昼夜循环等功能设定，两者生成时间都是15分钟，在细节效果不相上下的情况下，不少网友表示更喜欢 Kimi K3.1 的整体视觉效果。不过，在一项 "哪家模型能追上 Claude Opus 5.5" 的实测中，Kimi K3.1 暴露了明显短板：它生成的长颈鹿奔跑画面，动作不符合正常的生物运动轨迹，观感十分怪异。总体来看，Kimi K3.1 已经证明了中国大模型在 "代码 + 视觉 + 推理（Reasoning）" 的多模态交叉领域，具备了不输于美国顶尖闭源模型、甚至在特定实用场景下超越它们的实力。
这有望大幅降低各行业开发 3D 与 Agent 应用的门槛，对于开发者做创意应用、企业做可视化项目、团队做自动化工作流而言，都是福音。Kimi 的视觉能力之所以这么猛，核心是它走了一条和传统文生图完全不同的路。简单说，传统模型是 "凭印象画"，容易出现结构崩坏、物体穿模；Kimi 是先写严谨的 3D 图形学代码，再渲染出来自查，发现 "轮子扁了"" 颜色不对 " 就自动重写代码修正，相当于边画边验算。这套能力不是一蹴而就的，Kimi 已经布局了近两年： 2024 年 12 月，Kimi 推出 K1 视觉思考模型，首次把强化学习引入视觉领域，在物理、化学图形题盲测中首次超越国外标杆； 2025 年 4 月，Kimi 开源 Kimi-VL 系列模型，基于 MoonViT 架构，仅 28 亿激活参数，就在 MathVision 数学视觉和 OCR 领域打平百亿级大模型，解决了 “看图费算力” 的痛点； 2026 年 7 月，2.8 万亿参数的 Kimi K3 发布，直接拿下 Arena 前端代码竞技场全球第一，在数据可视化、交互动效等 6 大前端领域超越顶尖闭源模型，奠定了 Web 3D 与图形生成的统治力； 如今的 K3.1，从匿名测试的表现看，其在3D与前端场景的优势有望延续。一方面，Kimi 用不到十分之一的成本，在 3D 交互、前端工程这些高实用价值场景里，追平甚至局部反超了海外顶尖闭源模型。另一方面，它会重点提升 Agent 长时间连续执行任务的能力。为此，团队计划利用后期训练、电脑操作和真实任务数据，进一步优化模型。与此同时，据野村证券的报告显示，月之暗面正在研发一款参数量预计达 5 万亿至 6 万亿的更大模型（外界称为 Kimi K4）。相比今年 7 月发布的 Kimi K3，新模型的参数规模接近翻倍。但目前他们正面临算力不足的最大障碍，具体训练进度尚不清楚。参考链接：https://x.com/Kimi_Moonshot/status/2113445757185606062
上车，雷峰网(公众号：雷峰网)带你看遍全球 AI 顶会精华
可独家畅览：
专家演讲PPT
大会报告全文
热门论文解读
学术新星访谈
扫描上方二维码
或点击「阅读原文」关注专区。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
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。
15 Minutes of delivery of industrial class 3D interactive assets rewrite the "Scientific Law" of visual LLM
I'm the author.
Edit Peak
Just now, Kimi's official publish has a "very implied" pilot film, with high-energy scientific images rapidly switching from particle-specific effects, and eye-to-eye Internet friends have captured the egg "Kiki K3.1."
Kimi K3.1 is coming? This time it's coming with a strong visual effect, and some netizens say that the official video was made by Kimi K3.1.
Image: Eye-pointed netizens found a bright spot in the image of Kimi K3.1 AI blogger Chris, not only confirming the news of publish, but also indicating that he had measured it for a week. He revealed that Kimi K3.1 performed well on the open source 3D Graphic Library Three.js and that the model cost 70 per cent of Kimi K3 and that Token was fully upgraded. From Chris' real picture of the sun, it can be seen that Kimi K3.1 ran out of a 24-hour round-the-clock simulation of the environment, from the morning light to the nightlight, the light to the frame, and the details of the environment flowed in tandem with the light. You know, Three.js prepares programs that are very resource-consuming because it requires real-time calculations of peaks, photos and textures in 3D space. And Kimi, this wave is equivalent to running out of grade paints in pure code. In fact, more details have already come out of Kimi K3.1. At the end of September, a model logo for “Kimi-k3-1” was found by developers in the configuration of the API platform on the dark side of the moon. On October 21, Kimi Code's model transition list quietly added the K3.1 option, which is equivalent to an official advance small-scale test for developers. In late October, two mysterious models, code names Kibitz and Kalmia, emerged suddenly from the arena. Of which, Kibitz is strong in frontend and 3D code type assignments. In the replicating missions like "Pelican Bicycling", which test the imagination of space, it produces almost perfect 3D dynamic images and vector maps with excellent geometry and physical instincts; in other 3D vixen (Voxel) rendering blinding measurements, it wins with precision control over super-long commands. It is worth noting that when producing 3D codes of equal complexity, it consumes only one tenth of the top foreign model. Soon, a developer passed the terminograph characterization test, assuming that Kibitz was Kimi K3.1. This is fully consistent with the pre-Kimi naming pattern, K2.5 is Kiwido, K2.6 is Kiwire, K3 is Kivine, K3.1 corresponds to Kibitz. Subsequently, a consolidation request from the SGLang code library unexpectedly revealed more details: K3.1 added a "Medium medium thinking" rating, which was expanded from three to four, allowing flexibility in the movement of calculations to accommodate speed and depth depending on the difficulty of the task. In Arena's blind eye, Kimi K3.1 inherited and further amplified the advantages of the front end of the web page and the graphic rendering, showing some degree of dominance in the 6 major subdivisions of 3D graphics (Three.js/WebGL), interactive screens, H5 dynamic simulations, etc. There's already a lot of cases coming out of the web, and the netizens are saying, "This is a 3D beast." To test the visual and spatial reasoning of LLM, the AI Circle has an inconvenient test "Pelican Bicycling." The subject appears to be simple, and the reality is that pits are everywhere. Since this picture does not refer to the ready-made pelican asset, it is important to understand the anatomical structure of pelicans and the mechanical logic of bicycles at the same time, and to combine two completely irrelevant concepts to calculate geometry in pure code. This will test the model's “concept combination capability”, “physical intuition capability”, and space geometry capabilities. Early on, Claude 3.5 drew herself. The car is just two circles and a triangle, and pelicans are a ball with a banana mouth, which is called "the abstract representative." And the results of Kimi K3.1 show that the difficulty is simply being filled. Figure: The early Claude 3.5 picture of the "Pelican Bicycling" shows that Kimi K3.1 made a hyena of cycling pelicans that conformed to a certain physical pattern, with a heavy weight and a steady body. With the exception of pelicans, where the moon hangs in the night sky, where the long contours of the remote hills line up to the orange red sunset sunside, the light beams are fired from the left red and white stripes and the background huts are glowing. In detail, the waves, the crabs, the long-distance sailboats cut the whole thing. The entire picture is complete, the spectrum of light is rich and the space relationship of all objects is perfectly correct. And more importantly, this is not a picture of pixels, it's a vector rendering from a line of code, like a mathematical formula that perfects the whole fairy tale scene. In response, the netizens stated: “This stacking of maps + photo-levels + object relationships understand that SVG is no longer a simple level that can be manipulated at random, that completes directly against the shoulder Opus 5.5, and the front-end visual capability of the national production model has been killed into the first ladder.” A netizen directly created a medieval castle with Kimi K3.1, and it was amazing. The details of the castle are very rich: the main object is a layered multi-storey tower with a flag on the top of the tower, which rises slightly in the wind, and the gate between the moats, surrounded by mountains, grasslands and wrong houses. There are also birds and clouds in the sky, and the light is pressed into the twilight, and the last light is warmed, and the wall of stone forms a clear and dark transition, as if the whole castle was "sustained in the air at the end of the day." Even more difficult is the fact that suspended rock bodies, piled towers, attached steps and vegetation are all in the same spatial logic and are in harmony with each other. However, after testing, the netizens said, "It feels like GPT-61 Sol Max." It was also noted that Kimi K3.1 added high-level controls, time control, rendering quality switching, etc., which would enrich the scenery. Another case is the generation of a whole steam train that can run, and Kimi K3.1 not only delivers very quickly, but is also highly complete. In particular, it can "discrete" the entire train by showing you: the head of the car, the car, the black cylinder gas tank, the red wheel, all parts of which have been disassembled into stand-alone parts, suspended in error in the image, each of which is complete and precise, like a moving "engineering chart." From the perspective of 3D interactive engineering, Kimi K3.1 found a new power breakthrough. Most of the mainstream models abroad remain in static modelling and basic image visualization, while Kimi K3.1 combines structural design, mechanical connection, interactive logic and complete disassembly view, dynamic velocity, real-time data tracking. These description models understand the principles of physical engineering rather than just a pile of vision. It retains both the GPT-style common understanding and the Anthropic-type deep logic, which may be a unique path for the national production of AI technology. The high-precision, interactive 3D helicopter model is completed in 15 minutes with a network member using Kimi K3.1. Those who know how to do so state that these are valuable 3D assets that can be significantly reduced to the prototype production cycle. Given Kimi K3.1's excellent visual performance, netizens can't help but compare it to the world's top visual model. In contrast to GPT-6 Astra, both models are built in My World, a robotic building, and a spherical building, each of them seems to be. Chiaki, but Kimi K3.1 costs only about 1/14 GPT-6 Astra. Figure: GPT-6 Astra and Kimi K3.1, in comparison to Haiku 5.5, both models are based on a model of a village, which includes the functions of dragon attacks, thunderstorms and day and night cycle, both of which are 15 minutes in length, and many netizens have expressed a preference for Kimi K3.1 ' s overall visual effects, given that the details are not comparable. However, Kimi K3.1 exposed the obvious shortboard in a test of which model can catch Claude Opus 5.5: It produces giraffe run-off images that do not fit the normal biological trajectory and are very strange. On the whole, Kimi K3.1 has proven that China’s LLM has acquired the power of the US’s top closed-source model in the multi-modular intersection of “code + visual + reasoning.”
This is expected to significantly lower the threshold for the development of 3D and Agent applications across industries, and is a blessing for developers to do creative applications, for businesses to do visual projects, and for teams to do automate workflow. Kimi's visual ability is so strong that at its core it's taking a completely different route from the traditional text-to-image. In short, the traditional model is an "impressive painting" that is prone to structural breakdowns and object dressing; Kimi is a first-hand 3D graphic code, reproducing from the search, finding that the "wheel flat" color is not the right color, and rewriting the code automatically, which is the equivalent of a drawing-by-worker. In April 2025, Kimi Open Source Kimi-VL Model, based on the MoonViT structure, was only $2.8 billion active parameters, in the MathVision Mathematical Visual Visualization and OCR Fields, to address the pain of “seeing the cost calculation” of K1 Visual Thinking Models, for the first time introducing intensive learning into the visual field, and exceeding the external pole for the first time in a physical, chemical graphic blinding; Kimi Open Source Kimi-VL Series Model, based on the MoonViT structure, with only $2.8 billion active parameters, in the MathVision and OCR Fields, to overcome the top-end closed-source model by setting up Web 3D with graphic power; K3.1 of July 2026, with 2.8 trillion parameters, with the expected continuation of its performance from the 3rd parallel to the 3rd point of the anonymous testing. On the one hand, Kimi, at less than a tenth of the cost, in the high-value scenario of 3D interaction, front-end engineering, even partially flattened the top closed-source model overseas. On the other hand, it will focus on enhancing Agent ' s ability to carry out successive missions over time. To this end, the team plans to further optimize the model using post-production training, computer operations and real task data. At the same time, reports from Nomura securities indicate that the dark side of the moon is developing a parameter count that is expected to be between 5 trillion and 6 trillion more LLM (externally known as Kimi K4). Compared to Kimi K3 of publish this July, the size of the new model is nearly double. At present, however, they are facing the greatest obstacles to their ability to calculate, and the specific training schedule is not yet clear. Reference link: https://x.com/Kiki Moonshot/status/213445571856062
Get in, Ray Peak.
It's an exclusive view:
Expert presentation PPT
General Assembly
The reading of popular papers
Academic New Star Interview
Scan 2D above
This post is part of our special coverage Syria Protests 2011.
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.