K3.1 Pro 的版本号只往前挪了 0.1，月之暗面想回答的问题却大得多。
作者｜宇航猿
编辑｜靖宇
10 月 23 日上午，Kimi 官方在 X 和视频号上发了一段 24 秒的视频。
画面里是一台复古风格的老式电脑，屏幕上开着 Kimi Code，终端里一行行地写着 Three.js 代码，在生成一个复杂的 3D 场景。眼尖的网友很快在左上角的模型栏里找到了答案——「Model: Kimi K3.1」。
Kimi 官方发布的 K3.1 预热视频｜图片来源：X @Kimi_Moonshot
用一台老古董般的机器，跑一段今天最前沿的 3D 代码，这个反差本身就是月之暗面想说的话。
十几个小时后，10 月 24 日零点刚过，卡着程序员节，Kimi K3.1 Pro 正式上线，同步推出的还有一个速度快 2 到 3 倍的 Highspeed 版本。官方承诺，完整模型权重将在 11 月 4 日前放出。
距离上一代 K3 发布，过去了三个多月。这三个月里，月之暗面经历的事情，比大多数公司一年还多。
01
「加量不加价」的 0.1
先说清楚 K3.1 Pro 是什么。
它不是一个全新的模型。月之暗面明确表示，K3.1 Pro 是 K3 的后训练版本，底座没换——仍然是 2.8 万亿总参数、1040 亿激活参数的混合专家架构，仍然是 100 万 token 的上下文窗口，仍然原生支持图像和视频输入。
变的是后训练。官方的说法是投入了「大幅增加的后训练算力」，在数百个下游领域做了大规模联合强化学习，为此把强化学习框架、沙箱基础设施、长上下文推理这些底层工程都重构了一遍。
也就是时候，同一个大脑，被重新「教」了一遍怎么干活。
效果体现在榜单上。在考察编程智能体的 Terminal-Bench 4.0 上，K3.1 Pro 拿到 65.7%，超过 GPT-6 Astra 的 58.2%，和 Claude Opus 5.5 的 66.4% 只差不到一个点。上一代 K3 在这个榜单上的成绩，按第三方机构 Artificial Analysis 的测法只有 13%。
第三方的综合评分也给了面子。Artificial Analysis 智能指数上，K3.1 Pro 拿到 55 分，比 K3 高出 11 分，排名全球第四，前面只有 Claude Opus 5.5、Claude Sonnet 5.5 和 Muse Wonder Full。等权重正式放出，它将成为得分最高的开放权重模型，开源与闭源之间的差距会从 12 分缩小到 3 分。
K3.1 Pro 在 14 项编程、科学与知识工作评测上的表现｜图片来源：Kimi 官方博客
除了编程，这次的两条新主线是 3D 和金融。
K3.1 Pro 能驱动 Unity、Unreal、Godot 这样的游戏引擎，能写 Blender 和 CAD 脚本做参数化建模。在从多视角图片重建 3D 模型的 BenchCAD 测试里，它借助工具拿到 92.7%，接近 Opus 5.5 和 Astra。Kimi 网页版和 App 也同步上线了「3D Make」，可以直接描述需求，生成 3D 网页、3D 模型，甚至导出给 3D 打印机。
kimi.com 上的 3D Make 界面｜图片来源：Kimi 官方博客
回头再看那段预热视频，老电脑屏幕上跑的 Three.js，其实就是在给这条主线打预告。
金融方面，K3.1 Pro 在基于真实财报做多步数值推理的 OfficeQA Pro v2 上拿到 64.4 分，官方称是目前最好成绩。在官方自己设计的高难度幻觉测试里，事实性错误率从 K3 的 23.4% 降到了 8.3%——当然，这是一个刻意找茬的压力测试，不代表日常使用的错误率。
金融类评测成绩｜图片来源：Kimi 官方博客
最让人意外的是价格。
K3.1 Pro 的 API 定价是每百万 token 输入 13 元、输出 68 元，比 K3 便宜了三分之一，大约是 Claude Opus 5.5 的一半。更强的新模型反而比旧模型便宜，这在大模型行业并不常见。
不过 Artificial Analysis 也算了一笔账，K3.1 Pro 完成同样任务要多用 14% 的输出 token，所以每个任务的实际成本大约 1.98 美元，和 K3 基本持平。
02
真实世界里的 K3.1
榜单之外，网友的第一手反馈要复杂得多。
好评集中在「干活」上。知乎上有开发者拿自己一个 Rust、C、汇编、CMake 混合的项目测试，认为 K3.1 Pro 的代码能力明显强过 DeepSeek V4.1 Flash，尤其对 Rust 常见设计模式掌握得更好，还顺手修掉了 DeepSeek 前几天留下的问题。同样的任务，K3.1 Pro 花了 20 分钟，换成 K3 估计要 40 分钟到 1 个小时。
速度是另一个被反复提到的改进。K3 的慢曾经是出了名的，有人形容慢到「可以流畅阅读它的英文思考过程」。到了 K3.1 Pro，思考过程大多一闪而过，来不及看就被折叠了。
但前端审美成了争议焦点。
有论坛用户用自己的私有题目做了一轮单文件 HTML 测试：云上龙城、琉璃机械孔雀、钻石。结论相当不客气，认为 K3.1 Pro 在已测模型里排在倒数，孔雀羽毛的高光甚至被画到了羽毛背后。另一位开发者让它根据参考照片用 Three.js 还原一个 Steam 手柄，改了三轮，LOGO 和部件的空间关系还是不对。
同一提示词「琉璃机械孔雀」，K3.1 Pro（Max 推理）的产出｜图片来源：大佬说论坛网友 PoemFarAway
同一提示词下 GPT-6.1 Sol（xhigh 推理）的产出｜图片来源：大佬说论坛网友 PoemFarAway
Steam 手柄还原测试提供的参考图片｜图片来源：大佬说论坛网友 NightRain
K3.1 Pro 第二版产出，LOGO 和部件关系仍有问题｜图片来源：大佬说论坛网友 NightRain
一边是官方 3D 演示的惊艳，一边是网友「买家秀」的翻车，两种声音都只是个案，但足以说明 K3.1 Pro 还没有到「闭眼用」的程度。
而最集中的怨言，指向的不是模型，是套餐。
有用户反映，在 199 元的老套餐里，让 K3.1 Pro 审查一个小游戏就烧掉了 5 小时额度的 13%，同样的任务交给 Opus 只用了 3%。还有人说 Highspeed 版用起来「真的很舒服」，但 5 小时的额度撑不到 2 小时。API 降价了三分之一，订阅用户却没有明显感受到。
测试进行到第三版时，199 元套餐触发 5 小时用量上限｜图片来源：大佬说论坛网友 NightRain
这其实是 K3 时代就留下的老问题。7 月 K3 发布后，因为请求量远超预估，Kimi 一度暂停了 C 端新用户订阅。算力，始终是悬在这家公司头上的那把剑。
03
回应指控的另一种方式
K3.1 Pro 发布的时间点，很难不让人联想到过去几个月围绕 Kimi 的那场风暴。
今年 2 月，Anthropic 公开指控 DeepSeek、月之暗面和 MiniMax 通过大量虚假账号与 Claude 交互、提取能力，其中归到月之暗面名下的交互超过 340 万次。
7 月 K3 发布、性能直逼 Claude Fable 5 后，白宫科技政策办公室主任克拉齐奥斯直接在 X 上发文，称月之暗面在开发 K3 时「蒸馏」了 Anthropic 的 Fable 模型，美国财长贝森特随后表示可能对其实施制裁、列入实体清单。
9 月，Anthropic 又发布了一份威胁情报报告，称月之暗面通过 5380 个伪造账户组成的代理网络，在 10 天里把近 30 万条 Kimi 用户请求转发给 Claude，并从中提取思维链用于训练。
对这些指控，月之暗面的回应一直很克制。
7 月，Kimi 企业业务负责人黄震昕表示，K3 的性能跃升来自底层原创架构创新，并非蒸馏复刻；7 月底，K3 开放权重的同时公开了技术报告，把混合线性注意力、注意力残差这些自研架构摊开给所有人看。
所以 K3.1 Pro 可以被看作一次不说话的回应。
最有力的反驳从来不是声明，而是一个对手没法假装看不见的分数。
这次 K3.1 Pro 最亮眼的几个数字，不少来自 Artificial Analysis 这样的第三方，而不全是自家跑分。更值得玩味的是，按 Anthropic 报告里的说法，Claude 现在已经不再对外返回完整的思维链，蒸馏的一方「拿到的是门牌号，不是钥匙」。在这样的封锁之下，K3.1 Pro 依然在三个月里把智能指数拉高了 11 分，至少说明月之暗面自己的后训练体系是真能出活的。
还有一个容易被忽略的细节。月之暗面在这次发布中主动披露，为了不让开源权重把网络攻击能力的前沿往前推，他们在训练中剔除了网络安全相关的数据。代价是在内部漏洞利用测试中，成绩从 90.5% 降到了 81%。
内部漏洞利用测试 xdev-bench-hard 成绩，K3.1 Pro 主动做了减法｜图片来源：Kimi 官方博客
在被扣上「安全威胁」帽子的当口，主动给自己的开源模型做减法，这本身就是一种表态。
但也要承认，分数能回答「你行不行」，回答不了「你从哪来」。
K3.1 Pro 发布当天，就有网友指出它画的鹈鹕 SVG 和 GPT-6 Astra 的风格几乎一模一样，怀疑对象从 Anthropic 换成了 OpenAI。这当然只是论坛上的猜测，一只鹈鹕证明不了什么。但它提醒我们，只要训练数据的来源无法被外界审计，「蒸馏」这顶帽子就很难被一个分数彻底摘掉。
11 月 4 日的权重发布，会是下一个观察点。权重开放后，全世界的研究者都能拿它去做测试、做对比、做分析，这比任何声明都更透明。
那台老电脑上跑着的 3D 代码，月之暗面已经给出了答案。至于更难的那道题，答卷才刚刚交上去。
*头图来源：月之暗面
本文为极客公园原创文章，转载请联系极客君微信 geekparkGO
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The version number of K3.1 Pro has been moved forward only to 0.1 and the dark side of the month is much larger.
Author of the Astronomical Apes
Editor
October 23rd morning, Kimi officially posted a 24-second video on X and video.
It's an old, retro-style computer, with Kimi Code on the screen, a line in the terminal with the Three.js code, creating a complex 3D scene. The top-eyed netizens soon found the answer in the top-left model bar, "Model: Kimi K3.1".
Kimi, official publish K3.1, pre-heat video photo source: X@Ki Moonshot
Using an old machine to run a 3D code that's at the front of the day, this contrast is what the dark side of the moon wants to say.
Ten hours later, on October 24, just past zero, the programmer's day is stuck, Kimi K3.1 Pro is officially on line, and there's another 2 to 3 times faster version of Highspeed synchronized. The official commitment is that the full model weight will be released by 4 November.
More than three months have passed since the last generation of K3 publish. In these three months, the dark side of the moon has experienced more than a year in most companies.
01
"Accountable" 0.1
Let's start with K3.1 Pro.
It's not a completely new model. The dark side of the moon makes it clear that K3.1 Pro is a post-training version of K3 and that the bottom seat is not changed - it is still a hybrid specialist structure of 2.8 trillions of total parameters, 104 billion of activated parameters, and context window of 1 million token, which still supports images and video input.
Changed to later training. Officially, "a much larger post-training calculus" has been put in place, and large-scale joint intensive learning has been done in hundreds of downstream areas, which has resulted in the re-engineering of these ground-level projects, namely, the enhanced learning framework, the sandbox infrastructure, and the long context reasoning.
And that's when the same brain was "teaching" again how to work.
The effect is on the list. K3.1 Pro received 65.7 per cent of the survey programming AI agent, 58.2 per cent more than GPT-6 Astra and 66.4 per cent less than one point less than Claude Opus 5.5. The results of the previous generation K3 on this list are measured by the third-party agency Artificial Analysis at 13%.
The third party's combined score was also a credit. On the Artificial Analysis smart index, K3.1 Pro received 55 points, 11 points higher than K3, ranked fourth in the world, with Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full ahead. When the weight is officially released, it will become the highest score open weight model and the gap between open and closed sources will be reduced from 12 to 3.
K3.1 Pro pictogram of performance in 14 programming, science and knowledge work evaluations: Kimi official blog
Besides programming, the two new main lines this time are 3D and finance.
K3.1 Pro drives a game engine like Unity, Unreal, Godot, and can write Blender and CAD scripts for parametric modelling. In the BenchCAD test, which reconstructs the 3D model from multiple perspectives, it obtains 92.7% with tools, close to Opus 5.5 and Astra. Kimi web version and App are also online "3D Make" and can directly describe demand, generate 3D web pages, 3D models and even export to 3D printers.
Kimi.com 3D Make Interface Image Source: Kimi Official Blog
Turning back to the pre-heat video, the Three.js that ran on the old computer screen, was actually warning the main line.
On the financial front, K3.1 Pro received 64.4 points on the OfficeQA Pro v2 multi-step numeric reasoning based on true financial statements, which is officially the best for the time being. In the government's own highly difficult hallucination tests, the factual error rate dropped from 23.4% to 8.3% in K3 - Of course, this is a deliberate stress test, which does not represent a daily error rate.
Source: Kimi Official Blog.
The most surprising is the price.
K3.1 Pro's API pricing is $13 per million token input, $68 output, one third cheaper than K3, about half of Claude Opus 5.5. Instead, the new, stronger model is cheaper than the old model, which is not common in the LLM industry.
However, Artificial Analysis also makes an account and K3.1 Pro uses 14% of the output token to do the same job, so the actual cost per mission is approximately US$1.98, roughly the same as K3.
02
K3.1 in the real world.
Beyond the list, first-hand feedback from netizens is much more complicated.
It's good to focus on "work." Zhihu has developed a mixed project of its own, Rust, C, Compiled, CMake, which considers K3.1 Pro to have a significantly stronger code capability than DeepSeek V4.1 Flash, especially with regard to Rust's common design model, and has been able to repair DeepSeek's problems over the last few days. The same mission, K3.1 Pro, took 20 minutes to replace it with K3 estimated at 40 minutes to 1 hour.
Speed is another improvement repeatedly mentioned. K3 was once famous for its slowness, which has been described as "the process of thinking in English that can be read fluently." At K3.1 Pro, most of the thinking is flashy, and it's folded before it's too late.
But the front-end aesthetic became the focus of controversy.
A single-document HTML test was conducted by Forum users using their own private title: Cloud Dragon City, Luo-ryung Peacock, Diamonds. The conclusion is quite generous, that K3.1 Pro is in the bottom of the measured model, and that peacock feathers are even painted behind them. Another developer made it use Three.js to restore a Steam handle based on a reference photo and changed the space relationship between the LoGO and the parts in three rounds.
The same prompt, K3.1 Pro (Max reasoning) output. Image by Big Brother Poem FarAway.
The same prompt GPT-61 Sol (xheigh reasoning) output photo source: Big Boss says forum member Poem FarAway
Steam, a reference photo source from the handle restoration test: Big brother said Forum NetRain
K3.1 Pro 2 output, LOGO and widgets still have a problem picture source: Big Boss says Forum NetRain
On the one hand, the official 3D demonstration was amazing, and on the other hand, there was a "buy-a-home" overturn, both of which were just a case in point, but were sufficient to show that K3.1 Pro had not reached the point of "closed eyes."
And the most focused grievances are not about models, but about a set.
According to the user, in an old $199, K3.1 Pro was asked to review a small game and burned 13% of quota for 5 hours, the same task was given to Opus for 3%. Others say that the Highspeed version is "very comfortable" but that five hours of quota won't last two hours. API has fallen by one third, and subscribers have not felt it clearly.
When testing to Release 3, the 199-dollar package triggers a 5-hour maximum. Image by Big Boss.
This is actually an old problem from the K3 era. In July, after K3 publish, Kimi temporarily suspended a new C-end subscriber because the request was far above estimated. Calculator, always the sword hanging over this company.
03
Another way to respond to allegations
K3.1 Pro publish points, it's hard not to think of the storm surrounding Kimi in the last few months.
In February this year, Anthropic publicly accused DeepSeek, the Dark Side of the Moon and Mini Max of interacting with Claude through a large number of false account numbers and extracting capabilities, more than 3.4 million times under the dark name of the Moon.
In July, after K3 publish, who was able to force Claude Fable 5, the Director of the White House Office of Science and Technology Policy, Kratzhos, wrote directly on X, stating that the dark side of the moon had "distilled" the Anthropic Fable model during the development of K3, and that United States Treasury Secretary Besent had subsequently indicated that it might be sanctioned and placed on the physical list.
In September, Anthropic also sent a threat intelligence report, publish, stating that the dark side of the month passed through a network of 5380 forged accounts, transmitting nearly 300,000 Kimi user requests over 10 days to Claude and extracting them from them for training.
The response to these allegations on the dark side of the month has been very restrained.
In July, the chief of Kimi ' s business, Huang Shui, stated that K3 ' s performance had evolved from a bottom-up original structure innovation rather than a retort; and at the end of July, K3 ' s open weight was accompanied by a technical report that spreads the mixed linear attention, the attention deficit, to all.
So K3.1 Pro can be seen as a silent response.
The strongest rebuttal was never a statement, but an opponent could not pretend to be invisible.
K3.1 Pro's brightest numbers this time, many of them from third parties like Artificial Analysis, not all home runs. More interestingly, as stated in the Anthropic report, Claude is now no longer returning to the full mental chain, and the distilled side "is getting the door plate, not the key." Under such a blockade, K3.1 Pro has raised the intelligence index by 11 points in three months, at least to show that the dark side of the moon is actually living with its own system of post-training.
There is one more detail that can easily be ignored. The dark side of the moon was voluntarily disclosed in publish this time, and in order not to allow open-source weights to push forward the front line of cyberattack capability, they were removed from the training for cybersecurity-related data. The price was a drop from 90.5 per cent to 81 per cent in the internal loophole utilization test.
Internal loophole test xdev-bench-hard performance, K3.1 Pro has taken the initiative to take a photo of the reduction: Kimi Official Blog
It is in itself an expression of the initiative to reduce its own open source model in the face of a "security threat" cap.
But also admit that the points answer "Can you do it" and "Where are you from?"
K3.1 Pro publish, on the same day, an online user pointed out that it painted a pelican SVG and GPT-6 Astra in almost the same way, with the suspect changing from Anthropic to Open AI. Of course, it's just speculation in the forum that a pelican proves nothing. But it reminds us that as long as the source of the training data cannot be audited by the outside world, the hat "distilled" is hard to get completely removed by a single score.
The weight of publish on November 4th will be the next observation point. When the weights are opened, researchers around the world can test, compare and analyse them, which is more transparent than any statement.
The 3D code running on that old computer, the dark side of the moon has given the answer. As for the more difficult subject, the responses had just been submitted.
* Headline source: Dark side of the moon
This article is an original article from Polar Park.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.