终于，K3.1 Pro 更新了，我也等了好久了。
又到了三个月一次的“吹牛逼”时刻！
Claude、GPT、Gemini 狂轰滥炸一个月之后，Kimi 终于“挺身而出”！
我看了官方的资料，也看了很多例子，已经眼花缭乱了，“看起来”是真的很猛，很哇塞！
看的我都心动死了，昨晚都没睡好，不是说好了让我先玩的么？
是怕我下手重捏坏了么？
每次新模型发布，都有人专门找 Good Case 大惊小怪一阵子！ 我就专门找 Bad Case 哈其实我不是故意的！
我就是随手一测！
我测试是为了方便我自己选模型，顺便和大家分享一下。
努力做到不浪费钱，不浪费时间，不浪费感情，三不原则！ 怎么感觉有点像……
有些朋友喜欢先看结论，我就先说一句：相比 K3 而言提升明显，“又快又便宜”相比Opus5.5...我们选择不比 ！ 具体能力……请看正文！
我按节奏来，先看卖家秀，再看实测结果，然后再总结。
我会重点展示测试结果、时间消耗、tokens 消耗等方面。
过程中有发现什么特别的点，我也会单独写出来！
文章略长，眼睛累了，可以先点赞😄！
基础信息
我相信，这么热门的模型，所有细节立马就会传遍大江南北，然后添油加醋，越传越神！
各种数据，大家肯定或多或少看过了，不用我多说。
但是为了文章的完整性，为了照顾没有看过资料的朋友，还是把重点信息提一提！
核心信息看这个帖子就可以了：
看看这里有啥新东西。
100 万上下文，2.8T（很大）MoE，多模态，这都已经是老黄历了。
主要的新东西就是：3D 建模和游戏！
这也是当前的最大热点！我最近也测过很多，做过很多。
另外提到了金融工作，还有编码、代理、知识等方面。
还有一个高速版模型，这个别看了，一般人烧不起的。
不过，听说这个版本默认速度就挺快，能到 100t/s！
基准数据看这张图就可以用了：
这里写了很多基准啊，有些我不熟。
我就看第一个，大概排名是 Opus5.5、K3.1 Pro、GPT-6 Astra！
Kimi 大部分基准是 top3！
Kimi 你发这个基准考虑过 DeepSeek 和 GLM 的感受嘛？它们基本都是 down3 了！
基准一向都是仅供参考的！
另外再来看一个第三方榜单吧：
AA 拿下第四名，估计大家已经刷到很多次这张图了，哈哈！
我本来挺相信这个榜单的。但是 Muse 能排第三，我就不知道它是什么标准了。
性价比就看这个图：
这个图主要表达的意思是 K3.1 Pro，在这一群人中，很有性价比。DS 很靠左，但是很低！
这些数据谁看谁迷糊，就像拍照一样，在自己的镜头最帅，一到别人的镜头里，都原形毕露了。
其他天花乱坠、神乎其技的第三方评论，我一句都不想提，我已经看腻了！
我想大家也不怎么相信那些“影评人”了！
快速过完资料，马上要开始看测试结果了。
测试工具
关于测试工具，简单说两句！
Kimi 自家的 Kimi Code CLI 做得还不错，我之前 K3 的所有测试都是在上面完成的！
这次我会用界面版：
界面版可以直接同步之前的内容，有了目录，我查看和管理会比较方便一些。
从左边记录可以看到，我之前对 K3 进行过全方位的深入了解的。
我记得当时应该是跑了 60 亿的 tokens，很多例子还没来得及分享。
从右边可以看到，我已经选中了 K3.1 Pro，思考深度为 High 级别！
测试结果
终于，可以上结果了。整整测了一整天了！
为了方便参考，我们还是把常测的几个例子全都跑完了，然后加了两个新例子，方便测试游戏和 3D 能力。
赛博上河图！
这个例子，不用我多解释了，老粉都已熟得很了！
这个题目没有标准答案，但是很容易看出差异，需要模型有很强的能力，要融合赛博朋克和清明上河图的风格。
K3.1 Pro 的结果如下：
哈哈哈~~人船混行，相当行！
Kimi K3.1 Pro 的审美还是不错的，元素和细节也比较丰富了，但是空间感和对画面元素的理解还是不到位。
这个主要体现在：虹桥的方位不对，另外人行道和河面没有分清楚，全是水上漂！
然后我再给大家看一下 K3 的效果！
这么一比，K3.1 Pro 还是进步巨大的！
相比其他国产模型，也是确实属于第一梯队了！
如果要横向对比 GPT-6 Astra 和 Opus5.5，那就还差点意思。
它们的丰富度和空间感太强了，这两位都把虹桥横跨两岸的效果做出来了，各有各的特点。
天文机械表！
这个题目考验 SVG 绘画能力，也非常考验脑子。因为同时提出了一堆需求，大模型注意力不够，就会丢三落四。
结果如下：
这个样式还是很棒的，整个画面布局和结构都很好，核心元素也都在了，没缺指针，运转良好。
国产模型中应该顶级了！
只是这个配色和样式总感觉 G 里 G 气的！
不会是和 DS 一起拜师学艺学了蒸功夫吧！
再给大家看下 K3 的效果：
无论是按钮的布局，手表的表带，还是配色，应该都是新版本更好。
新版更加精致，更加有格调！
结果很好，但是过程中遇到了一些小问题：
它只是在屏幕中输出了代码，但是没有保存到文件夹中！
我的提示词中包含了：“只输出最终代码，不要解释。 不读取和修改当前目录的文件，单独保存一个页面”
它可能误解了我的意思，我的本意是不要参考已经存在的文件，单独保存一个页面。它理解为不要在这里创建文件，只要在屏幕中输出即可。
我后面明确说了，要单独保存一个页面！ 这一句也看不懂么？
我对它的理解能力表示怀疑！
这么干活，要扣工资的。
游戏复刻
审美、空间感、脑力都考完了！ 来玩会儿《超级玛丽》！
结果如下：
刚开始这一幕还算正常吧。玛丽老哥也算是眉清目秀。
长大之后就彻底长残咯：
我大致玩了一下，可以通关的，没有大问题！
但是角色、地图、玩法的还原度不够。
人物形象和地图风格大家已经能看到了。
玩法部分有几个明显的问题：它没有隐藏蘑菇、长大之后跳不高、按“X”加速太没人性了。我的手指🤌要怎么按啊。
我不吐槽，我不吐槽，我正在写一个正经的测评！
好不好，比较很重要，看看 K3 的：
这么一比，K3.1 是不是好很多！
这个例子很神奇，即便是最近更新的一些国产模型，做得都不咋滴。
比如最近更新的 DSV4.1_1017 做完后跳跃都有问题，MiniMax 3.1 直接啥都不显示。
换个表述就是 K3.1 比最近更新的几个国产模型要好。
Claude 和 GPT 咱们就不比了，它们的 Sonnet 和 Sol 做的效果已经非常逼真了。
这是 GPT-6.1 Sol 的！
甲维斯的车库
这个例子是专门用来测试网页复刻和 3D 建模的。
也是这次 K3.1 Pro 的重点宣传方向，我看了很多炫爆了的建模和游戏效果。我是非常期待它能在这个例子上艳压群芳的。
但是我遇到了一个问题，我正在怀疑，我是不是用了盗版模型。
我跑的结果如下：
这已经是修改之后的结果了！
这里有几个问题：第一个左上角的卡片样式全乱了，第二个轮子不滚动，第三个车子结构混乱，第四个 Mark III 完全不像！
渲染素一点问题不大，建模不像这个问题，就比较难受。
这个是 K3 的：
这是 GPT-6 Astra 的：
好像 Astra“略胜一筹”！
K3.1 Pro 这个例子整体来说不能算太差，但是不符合我的期待。
我看了那么多牛逼的宣传视频之后，我对它的期待是，它会轻松刷新这一题的新高度，但是并没有。
四个常规问题测完了，还可以，但是没我想象中的出众！
游戏手柄测试！
最近，我搞了一个新的系列，专门用来测试国产 Flash 模型的。
比如 GLM 5.3F、DeepSeek 4.1F、MiMo v3.1F 等等。
没想到第一个题目就让所有 Flash 都翻车了，我感觉要把题目升级为 Pro 测试题了。
今天 Kimi 3.1“Pro”来了，刚好拿来测试一下。
这个题目是这样的：
帮我在网页上画PS手柄，然后支持真实的手柄接入，按相应的按钮，会做出对应的反应。要考虑到多手柄接入，我最多会接入四个手柄。核心要求：1、网页制作漂亮，匹配主题，有设计感2、手柄轮廓和样式要高度还原3、要考虑幽灵插槽的问题，能分清楚不同的手柄。
K3.1 Pro 的结果如下：
它这个手柄并不是写实风格，而是按抽象线条的方式来展现的。
不够逼真，但整体来说还是一个结构完整、形态基本正常的手柄。
L 和 R 按钮都标出来了，细节还可以！
重点是，它确实能正常接入 4 个手柄。这一点确实就是 Pro 级别的模型该有的水平了！
这个例子做得不错！
坦克世界
坦克大战这种经典小游戏已经不能满足当前的测试了，所以我们把游戏升级到了更大的战场——“坦克世界”！
坦克世界是走真实路线的坦克大战！
我不知道这算不算 3A 大作，反正是一个制作不小的游戏！
我们这个测试的目标就是复刻它的玩法和效果。
我的制作方法是：让 K3.1 Pro 自己去找参考图，自己总结游戏特点和核心玩法。
这个过程中遇到一个问题，它的搜索抓取能力较弱，基本上没抓到有用的图片。
后来是我手动补上了参考图，好像我也是多此一举~~因为它实际上也没参考个蛋出来！
我们来看结果吧！
第一版结果如下：
看到这里，你们应该知道，我选择封面图的理由了吧——货不对板啊！
因为第一版问题比较多，比如鼠标乱跑，镜头乱晃，转弯转反，瞄准很难，UI 太卡通等等。
所以就改了一阵子！
下面第二版：
好了一些，但是参考图的效果估计是打死都改不出来了！
这是 GPT-6 Astra 做的：
单论建模和场景效果，经过我调整的 Kimi 3.1 Pro V2 版其实已经要比 GPT-6 好一些了！
只是它们都没有达到我的要求，坦克世界不是这种卡通风格。
这里还有一个非常有意思的点，它们整体气质非常像：
无论是游戏里面，还是 UI，或者配色！
我第一眼看到，就感觉似曾相识，K3.1 Pro 真的“G 里 G 气的”！
这一波我们可以简单粗暴地说，K3.1 Pro 打平 GPT-6 Astra 了~哈哈！
我也算是有个“胶带”了！
这里不拿Claude Opus5对比，否则很拿收场。
3D 建模 Blender
官方重点宣传了 3D 建模和游戏，游戏上面做了一个，下面看看建模！
前几天马斯克用 Roadster 怒刷存在感，这个车子确实也非常帅气。所以我想把它融入到游戏里。
所以专门做了一个建模测试，今天我也让 K3.1 Pro 来做一做！
这个测试例子最棒的一点是它与众不同，它的轮廓非常有特色，所有模型都还没有学习过它的轮廓。
最终建模效果如下：
我让它渲染了 10 种场景：
整体来说，有鼻子有眼的，车子是车子，轮子是轮子。大致看起来不错，渲染效果也不错！
K3.1 Pro 在 Blender 中的建模能力应该是比网页中用 three.js 强很多。
但是，细节还是不行：
一旦拉近看，就没法看。
这个问题不光是它的问题，Opus5.5 也不行，GPT-6.1 也不行。
GPT-6.1 Sol 在这个项目中做得最好！
不能说完美，但是它已经考虑到了部件建模了，细节拉满！ 左下图是不能起飞的那款，右下图是可以起飞的那款！
虽然 K3.1 Pro 的卖点是建模和游戏。但是大家都通过参考图片建模的情况下，它并不出众，GPT-6.1 Sol 和 Astra 在 Blender 中建模还是很强的，Opus5.5 在网页中建模很强。
至于其他国产模型，我还没试过，大概率会很抽象，估计只能看出来是个车子，但是很难做到 Roadster 的外形线条。
今天已经搞了 7 个例子了，够够的了，该收了！
下面说另外一个大家都很关心的话题。
时间和 tokens 消耗
能力很重要，时间、钱更重要！
Kimi 套餐谁用谁知道，前戏还没完，就结束了！
那么这次 Kimi 3.1 Pro 的消耗怎么样呢？ 由于我的套餐比较特殊，不太好代表大家看消耗。
所以我就统计了一下 tokens，然后再算出价格。
同时也把不同任务的时间统计一下！
统计数据如下：
这么多例子，居然只消耗了 4000 万 tokens，77 块钱，难以置信啊！
整体的耗时也算是比较短的。
我之前用其他模型测超级玛丽，最长干到了 6 个小时，一两三个小时也很常见。
一堆国产 Flash，说是速度很快，一旦跑任务，就慢得要死。
K3.1 Pro 大部分例子都在 20 分钟内，这个算是很快了，3D 坦克这个稍微长一些花了 40 分钟。
这个没毛病，当时 Opus 5 也花了个把小时。
至少从时间消耗上来看，已经打败其他国产模型，赶上世界领先水平了！
这就是我开头说的“又快又便宜”的意思了。这是一个相对值，相对上一个版本，以及相对我的预期而言，算是又快又便宜了。
做个小总结
我知道你们其实比较喜欢看我开炮，因为网上吹牛、乱叫的太多了。
但是，我成熟了！
我们要多看人家的进步，努力和优点。
当然，缺点，该提的还是要提，这个是很关键的，否则还测个毛线。
优点：
相比 K3 提升比较明显，大部分例子都有明显提升。
让我意外的一点是，Kimi 3.1 Pro 的速度非常快，这是我很关注的一个点。
现在很多国产 Flash 一点都不 Flash，而 Kimi 这个 Pro 却非常 Flash。
单任务消耗应该比 K3 低很多！
3D 建模能力其实也不错啦，这个在坦克世界和 Blender 的例子里有体现！
这都是实实在在有用的东西。
缺点：
快了之后容易扯到蛋， 我怀疑为了快和省，做了脑切除。
先说一个具体的问题，Kimi Code 内置的 Research 能力（搜索和分析）比较弱，几个项目收集截图都失败了。
长程任务方面，相比国际领先的模型还是有欠缺，快了之后，感觉就不愿意做太多了。
另外自主性方面也不够，无法做到“许愿式编程”，一轮出成品！
当然这个要求是高了些，但是 Opus5.5 确实可以！
上限不够高，我本身的期待是它能和 Opus5.5 有来有回，实际上是一拳被 K·O！
文章中 Opus5.5 的结果我没有放太多，不放有不放的原因！
另外发现一个负优化，网页和 PPT 设计感和多样化能力好像降低了很多！
有种 GPT 5.x 化的趋势！
另外，网络上很多牛逼的例子，可能是创作者本身能力强。
如果大家一定要一个确切的排名或者榜单。
我觉得这个排名和我的想法比较接近（Muse 另议）：
从一个侧面来说，相比 K3 的惊艳，三个月憋出的 0.1，并没啥让人兴奋的点。更多的是一种迫不得已吧！ 而且这次所有榜单排名都没有特别高，那么实际情况还是要打点折扣的。
主要是国外的模型太强了，当大家以为国模越来越接近美模的时候，美模突然就踩油门了……！一看推进器是火箭推进器~~
声明一下：
我的测试并不代表任何东西，只代表我的测试，我找题目是比较刁钻的，主要是看它某些极限能力。所以即便是结果不理想，也不意味着它在常规任务中做的不好。结果好的，我就不用解释了是吧，它就是做得好啊！
整体来说：
只要“套餐配额充足”，我还是很愿意用这个模型的，能力不错又不磨蹭，在当前也是非常优秀的品质了！
另外，不管模型怎么样，大家都很努力，点个赞吧。
这里面涉及到的对比实很多，我只放了很少一部分，大家可以自行查看，通过网站上的展示进行像素级对比。 K3.1pro的例子我也会马上上传。
后续，我还是打算深入探索一下 K3.1 Pro，如果挖到好东西，我会跟大家分享，如果没有，我就不发了。
话说有 Pro，是不是应该有便宜的 Flash？
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。
Finally, K3.1 Pro updated, and I've been waiting a long time.
It's a three-month bragging time!
A month after Claude, GPT, Gemini, Kimi finally came forward!
I've looked at official information, and I've seen a lot of examples, and I've seen a lot of things.
I couldn't sleep last night. Didn't I have to play first?
Are you afraid I'm gonna break it?
Every time the new model publish is, someone's been looking for Good Case for a while, and I'm looking for Bad Case.
I'm just testing!
I'm testing to make it easier for me to choose the model myself and share it with you.
Try not to waste money, not time, not feelings, not principles!
Some friends like to look at the conclusions first, so I'll start by saying that, compared to K3, advancement is obvious, "fast and cheap," compared to Opus 5.5... we choose not to compare!
I'll do the rhythm, see the seller show, see the results, and then sum it up.
I will focus on testing results, time consumption, tokens consumption.
There's something special going on in the process, and I'll write it out myself!
It's a little long. I'm tired of my eyes.
Basic information
I believe that such a hot model will spread all the details to the north and south of the river, and then add vinegar and the more the gods!
I'm sure you've seen more or less of the data.
But for the sake of the integrity of the article, to take care of friends who have not seen it, the key information is mentioned!
Read the post on the core:
Look what's new here.
One million context, 2.8T (large) MoE, multi-model, are already old.
The main new thing is: 3D modelling and games!
It's also the biggest hotspot of the day. I've been doing a lot of tests lately.
Financial work was also mentioned, including coding, agency, knowledge, etc.
There's a high-speed model. Don't look at it. It's not gonna burn.
But I hear this version's going pretty fast.
The baseline data can be used in this image:
There's a lot of benchmarks here, some of which I don't know.
I'm looking at the first one, probably Opus 5.5, K3.1 Pro, GPT-6 Astra!
Kimi, most of the benchmarks are top 3!
Kimi, have you sent this benchmark to consider how DeepSeek and GLM feel? They're basically down 3!
The benchmark is always for information!
Another third-party list:
AA, take the fourth. You've done this many times. Ha ha!
I kind of believed this list. But Muse can be third in line, and I don't know what it is anymore.
The price for money is this:
The main message of this figure is K3.1 Pro, which is very competitive among this group. DS is very left, but very low!
The data, who looks at who's lost, is like a photo, in his shot, the best-looking, when he gets to someone else's Z87Z.
I don't want to talk about it. I'm tired of it!
I don't think you believe in those critics anymore.
We're about to start reading the results.
Test Tool
A quick word about the test tools!
Kimi's own Kimi Code CLI did well, and all my K3 tests were done on it!
This time I'll use the interface version:
The interface version can be directly synchronized with what happened before. With the catalogue, it will be easier for me to view and manage.
As can be seen from the record on the left, I've had an in-depth understanding of K3 in all its aspects.
I think it was six billion tokens, and many examples were not shared.
As you can see from the right, I've selected K3.1 Pro, thinking at the High level!
Test Results
Finally, it's working. It's been all day!
For ease of reference, we ran out of the usual examples and added two new examples to facilitate the testing of games and 3D capabilities.
Saber up the river!
I don't have to explain this to you.
There is no standard answer to this topic, but it is easy to discern differences and requires a model with a strong ability to integrate the styles of Cyberpunk and the Qingming River Map.
K3.1 Pro results are as follows:
Ha-ha-ha-ha-ha-ha-ha!
Kimi K3.1 Pro's aesthetics are good, and the elements and details are rich, but the spatial sense and understanding of the graphic elements are still not in place.
This is mainly reflected in the fact that the bridge is not in the right direction, and the other sidewalks and the river are not clearly separated.
Then I'll show you the K3 effect!
So far, K3.1 Pro is still making great progress!
Compared to other national models, it's really the first one!
If we were going to cross-reference GPT-6 Astra and Opus 5.5, that would almost be true.
Their richness and spatial senses are too strong, and each of them has made a difference between the two sides of the bridge.
Astronomical meter!
This is a subject that tests the ability of SVG to paint, and it's also very brain-tested. LLM is not paying enough attention because of the demands made at the same time.
The results are as follows:
It's a great style. It's a great layout and structure. It's got the core elements. It's a good workout.
It should be the top of the national production model!
It's just this coloring and style feeling in G.
It's not like he's studying martial arts with DS.
Show you K3 effects:
The layout of the buttons, the watch belts and the colours should be a better version.
The new edition is more refined and more elegant!
The results were good, but small problems were encountered:
It just output code on screen, but not in the folder!
My prompt contains: "Only the final code, don't explain it. Do not read and modify files in the current directory, save a single page
It may have misunderstood what I meant by saving a single page without reference to an existing document. It understands not to create files here, just output in the screen.
I made it clear that I had to save a single page!
I doubt its understanding!
Work like this, it's a deduction.
Reset Game
The aesthetics, the sense of space, the brain!
The results are as follows:
It's normal to start with. Marion's kind of a lookout.
When I grew up, I was completely disabled.
I've had some fun. I can get through customs. No big problems!
But there is not enough reduction of roles, maps, games.
You can see the image and map style.
There are several obvious problems in the play: it does not hide mushrooms, it does not jump high when it grows up, and it is too inhuman to press "X" to accelerate. How do you press my fingertips?
I don't give a shit, I don't give a shit, I'm writing a serious test!
It's more important to see K3's:
So, K3.1 isn't much better!
This is an amazing example, and even some of the recently updated national production models are not doing much.
For example, the recently updated DSV 4.1 1017 has problems with jumping, and Mini Max 3.1 doesn't show anything directly.
In other words, K3.1 is better than a few recently updated national models.
Claude and GPT are not the same, and their Sonnet and Sol have done a very good job.
This is GPT-6.1 Sol!
Jarvis' garage.
This example is specifically designed to test web-page regraving and 3D modelling.
And this time K3.1 Pro focused on the promotion, and I saw a lot of modeling and game effects. I'm looking forward to it in this example.
But I had a problem, and I was wondering if I was using a pirate model.
I ran as follows:
It's already changed!
There's a few problems here: the first top left card style is all messed up, the second wheel is not rolling, the third car is in disarray, and the fourth Mark III is completely different!
It's not too much of a problem with retrofitting. Modelling doesn't look like it.
This is K3:
This is GPT-6 Astra:
It's like Astra "slightly wins."
K3.1 Pro can't be a bad example overall, but it doesn't fit my expectations.
After looking at so many bad propaganda videos, my expectation is that it will simply refresh the new height of the subject, but not.
Four routine questions are done. It's okay, but it's not as good as I thought!
Game handle test!
Recently, I've developed a new series for testing the national Flash model.
GLM 5.3F, DeepSeek 4.1F, MiMo v3.1F, etc.
I didn't think the first one would turn all Flash over, and I feel like I'm going to upgrade it to Pro.
Kimi 3.1 "Pro" is here today, just to be tested.
Here's the title:
Help me draw PS handles on the web page and then support real handle access, press the corresponding button and react accordingly. Taking into account multiple handles, I'll access up to four handles. Core requirements: 1; good web site production; matching themes; design sense 2; handle contours and styles to be reduced at a high level of 3; ghost slots to be considered; different handles to be identified.
K3.1 Pro results are as follows:
It's not written in real style, it's presented in abstract lines.
It is not true enough, but it is, overall, a well-structured and largely normal instrument.
The L and R buttons are marked, the details are fine!
The point is, it does have access to four handles. That's exactly the level of the Pro level model!
That's a good example!
Tank World
The classic game of tank warfare is no longer enough to meet the current test, so we have upgraded the game to a bigger battlefield, the Tank World!
Tank world is a tank battle on the real path!
I don't know if it's a 3-A game or not.
The goal of our test is to reset its play and its effects.
My way of doing it is to get K3.1 Pro to find a reference map for himself, to sum up the character and core of the game.
A problem was encountered in the process, where the search capability was weak and the useful pictures were largely missing.
And then I made it up manually, like I was doing it too, because it didn't actually look at an egg!
Let's see the results!
The results of the first edition are as follows:
See here, you know why I chose the cover sheet.
Because there's a lot of problems with the first version, like the mouse running, the shot swinging, the bending, the difficulty of targeting, the Ui too cartoons.
So I changed it for a while!
The following second edition:
Well, it's a little bit, but it's not gonna work out.
GPT-6 Astra did this:
The single modeling and scene effect, the Kimi 3.1 Pro V2 version of which I adjusted, is actually better than GPT-6!
It's just that they didn't meet my requirements. The tank world is not that cartoon style.
And here's another very interesting thing. They have a very similar whole.
Whether it's in the game, or the UI, or the color!
The first time I saw it, it felt like I knew each other, K3.1 Pro was really "G-G-G-G"!
This wave we can simply say, K3.1 Pro level GPT-6 Astra now!
I have a tape!
There's no comparison between Claude Opus 5, or it's over.
3D Modelling Blender
Officially, 3D Modelling and Game, one of the games, look at modelling!
The other day, Mask used Roadster's anger to draw a feeling, and the car was really handsome. So I want to put it in the game.
So I did a modeling test and today I let K3.1 Pro do it!
The best thing about this test example is that it's different, its contours are very distinctive, and all models have not learned its contours.
The final modelling effects are as follows:
I made 10 scenes of it:
Overall, there are noses and eyes, cars and wheels. It looks good in general, and it's a good rendering!
K3.1 Pro should have a much better modelling capability in Blender than three.js on the web page.
But the details are still not:
Once you get close, you can't.
It's not just about it, either Opus 5.5 or GPT-6.1.
GPT-6.1 Sol did the best in this project!
I can't say it's perfect, but it's got parts to model, the details are full!
Although K3.1 Pro's selling point is modelling and games. But it's not remarkable when you're looking at pictures, GPT-61 Sol and Astra are very strong in Blender and Opus 5.5 on the web.
As for other models of national production, I haven't tried yet, but it's probably abstract, and it's probably just a car, but it's hard to do the Roadster line.
It's been seven examples today. That's enough.
Next is another topic of concern to all.
Time and tokens consumption
Power is important, time, money is more important!
Kimi, who knows? The foreplay is over!
What about Kimi 3.1 Pro's consumption this time? Because I have a special package, it's not good to look at consumption.
So I counted tokens and then I figured out the price.
And count the different assignments!
The statistics are as follows:
With all these examples, only 4 million tokens, 77 dollars. Unbelievable!
The overall time-consuming is also relatively short.
I used other models to measure Super Marie, up to six hours, and it's common for a couple of hours.
A bunch of domestic Flash said it was fast, and it was too slow to run.
K3.1 Pro, most of the examples are in 20 minutes, which is pretty quick, and 3D tanks, which took a little 40 minutes.
That's okay. It took Opus 5 an hour.
At least in terms of time consumption, we've defeated other national production models to catch the world's lead!
That's what I said at the beginning: "Fast and cheap." It's a relative value, relative to the previous version and relative to my expectations.
Let's make a small summary.
I know you guys like to watch me shoot 'em, 'cause the Internet is bragging.
But I'm mature!
We need to see more progress, effort and merit.
Of course, there are shortcomings, and the point is still to be made. This is crucial, otherwise there is a hair line.
Advantages:
Compared to K3 promotions, the majority of cases are significantly higher.
The one thing that surprised me is that Kimi 3.1 Pro is very fast, and that's one of my concerns.
Now a lot of the state-owned Flash doesn't have any Flash, and Kimi this Pro is very Flash.
Single task consumption should be much lower than K3!
3D Modelling is actually good, as demonstrated in the example of Tank World and Blender!
These are really useful things.
Disadvantages:
Soon it'll be easy to get to an egg, and I suspect I've had a brain cut for speed and economy.
To begin with a specific question, Kimi Code's built-in Research capability (search and analysis) is weak, and several projects have failed to collect a screenshot.
In terms of long-range tasks, there is still a gap in the international lead model, and soon enough it feels reluctant to do too much.
There's not enough autonomy to be able to make a wish-formulation and produce a round!
Of course it's a little higher, but Opus 5.5 does!
The ceiling is not high enough, and my own expectation is that it can go back and forth with Opus 5.5, actually a punch by K.O.!
The result of Opus 5.5 in the article was that I didn't put too much in it for reasons!
And there's a negative optimisation, a web page and a PPT design sense and diversity capacity that seems to have been reduced a lot!
There's a GPT 5.x trend!
In addition, there are many examples of the power of the network, possibly the creators themselves.
If you have to have an exact ranking or a list.
I think this ranking is closer to what I think.
On one side, there's nothing exciting about K3, compared to K3, the three-month ounce. It's more of a necessity! And all the lists are not so high this time, so the reality is a little discount.
Mostly foreign models are too powerful, and when people think that Quino is getting closer to Quino, Quino suddenly stepped on the gas...! Looks like a rocket booster.
Declare:
My test doesn't mean anything. It's just my test, and I'm looking for something that's a little rough, mostly some of its limits. So even if the outcome is not ideal, it does not mean that it does not do well in its regular mandate. Well, I don't have to explain it.
Overall:
As long as the ration is adequate, I'd be happy to use the model.
Besides, no matter what the model is, we're all trying hard.
There's a lot of contrasts involved, and I've only put a tiny fraction of them, so you can look at them yourself and do pixel comparisons through the displays on the website. I'll upload the K3.1pro case immediately.
After that, I'm going to explore K3.1 Pro, and if I dig something good, I'll share it with you, and if I don't, I won't.
Pro, shouldn't there be cheap Flash?
Auto-fetched and translated by AI Atlas. All rights belong to the original author.