文 | 数智奔流，作者 | 田渝，编辑 | 沈校
10月24日凌晨，Kimi发布新模型K3.1 Pro。官方给它安排了两项主要升级：能力进入旗舰模型第一梯队，API价格较上一代下调约三分之一。
但模型发布后，kimi缺因为Token消耗速度陷入性价比争议。
从定价看，官方公布的API价格确实下降了。
K3.1 Pro的输入价格从每百万Token 20元降至13元，输出价格从100元降至68元。可在C端用户和海外开发者那里，实际感受并没有同步变便宜。
有用户测试经典的“鹈鹕骑单车”任务，生成时长达到1小时47分钟，5小时额度消耗87%，周额度消耗16%。还有开发者记录到，只发送一句“你好”，Token计数就达到4.2万。
10月26日，Kimi发布致歉和补偿声明，承认“问题出在我们，本应更早修复”。官方提出两项措施：为有效期内的订阅会员重置并补偿额度；调整模型配置，减少不必要的子Agent调用和输出长度。
这份声明承认，模型背后的调用机制没有被控制好。
过去，用户对模型价格的理解相对简单：输出多少Token，按照单价计算。
但现在，一次请求可能还包括长上下文读取、后台推理、任务拆解、工具调用、子Agent协作和多轮自我修正。
用户只发送一句话，模型可能在后台完成一串动作。只要每次任务消耗的Token更多，即使API降价，用户最终付出的成本或许还会增加。
K3.1 Pro被官方称为“迄今为止能力最强的3万亿级参数模型”，支持100万Token上下文，定位在长程编程、3D创作和金融研究，在这些优势场景下，本身就容易触发长上下文和多Agent调用的任务类型。
独立研究机构Artificial Analysis测试，智能指数上K3.1 Pro得分55，位列全球第四、开源模型第一，比K3提升了11分。按照这份榜单成绩，如果后续按计划开放模型权重，它也有望成为该榜单中得分最高的开放权重模型。
这些成绩说明，Kimi确实完成了一次能力跃迁。
但基模性能和产品体验之间，还有普通用户的真实使用体验。
为了观察K3.1 Pro在日常任务中的表现，我们使用Pro版订阅账号，通过Kimi电脑端应用连续测试了多项创作任务。
测试重点不是单次回答是否漂亮，而是四个环节：首次生成效果、修改过程、最终完成度，以及额度消耗。
截至10月27日下午，账户月度总使用量为27.03%，其中五项测试任务合计消耗约25.48%。
鹈鹕骑自行车几乎是每个新模型发布后都会出现的经典实测案例。它的源头，来自技术博主Simon Willison。原题只有一句话：Generate an SVG of a pelican riding a bicycle（生成一张“骑自行车的鹈鹕”的矢量图）。
SVG需要用代码描述图形，模型既要理解鹈鹕和自行车的结构，也要处理车轮、车架、鸟腿和踏板之间的空间关系。为什么Simon偏偏和“鹈鹕骑自行车”过不去？一是因为Simon本人喜欢，二是当时网上缺少同类素材，模型较难直接复现训练数据中的图片。
随着模型能力进步，这道测试逐渐成为新模型发布后的“开机仪式”。我们这次把题目升级为短视频，在结构和美观之外，继续观察骑行动作是否自然、画面是否连贯。
我们要求鹈鹕面带微笑，在夕阳下沿海岸线骑自行车，海浪不断拍岸。
任务一次完成，没有追加修改，最终效果不错，鹈鹕的表情生动，身上的毛发随风扬起；海浪拍打自然，整体画风温暖。
额度明细显示，K3.1 Pro模型条目消耗0.13%，Plugin条目消耗5.85%，合计5.98%。也就是说，这次任务没有返工，约97.8%的消耗来自Plugin条目。
接下来，我们让它制作一个苏式老院。
提示词是“做一个苏式老院，院子中有枇杷树，树下有藤长椅和桌子，桌子上放着枇杷和茶杯”。
第一版的成品比较粗糙，整体质感接近颗粒积木，树叶像纸片一样。
第二版，我们要求取消沙盘式俯视，换成院内自然透视与轻微景深；使用低角度暖阳，强化夏天氛围；把长椅换成躺椅，并继续完善其他细节。
这一版解决了视角、躺椅、蒲扇等主要问题，细节仍然存在偏差。庭院外的竹林过于简化，蒲扇悬空，枇杷还嵌进了树根。
第三版，我们提供了一张真实参考图，让它重新制作。有了参考图之后，画面内容更加丰富，问题集中在细节还原上。石柱位置突出，池塘周围的石头质感接近泡沫。
额度明细显示，苏式老院首次生成消耗1.76%，后续修改继续增加4.45个百分点，整个任务合计消耗6.21%，在五项测试中排名第二。
第三个任务是制作一个介绍世界美术史的网页。我们要求代表人物和画作足够丰富，用户点击人物头像或画作后，可以查看相关知识解答。
生成的网页名叫“ATHÈNE｜世界美术史·浪漫长廊”，以暖纸色、深棕色和柔和光影营造浪漫氛围，用户能筛选文化轨道、人物或画作，也能按章节漫游、搜索关键词，人物动效可唤醒。
第一版更接近一个精选画廊，只收录了13位人物和14件名作，数量偏少，中西艺术也混合在同一条线索里。点击“导览下一件”时，右侧文字解释和左侧画作存在不同步，画作大约比文字慢三秒才显示。
进一步优化后，网页内容扩充到30位人物和30件作品，以西方艺术、东亚艺术双轨呈现，并以时间轴清晰梳理脉络。点击任意人物头像、画作或时间轴节点，就能阅读相关背景介绍。手机端也能流畅浏览，细节兼顾美感、易用性。
美术史网页的模型条目合计消耗4.90%，Plugin条目消耗0.02%，总计4.92%。
接下来，我们重现了两位博主的测试。
第一个是机械键盘，我们要求它制作一个高精度模型，展示键帽、轴体、定位板、PCB、外壳等零件结构，同时支持白模、X-ray透视、爆炸视图和不同方向视角。
这次生成没有经历多轮重做，主要结构完整，多种视图也能正常切换。额度明细中出现了两条计费记录，分别为1.09%和0.37%，合计1.46%，是五项任务中最低的一项。
第二个任务是还原《桃花源记》。
第一版画面偏符号化，人物和物体由抽象元素替代，整体质感与苏式老院首版接近。我们随后提供了一张山水画参考图，让它重新制作。第二版仍然没有抓住原文的空间变化和叙事层次，画面与"桃花源"的意境存在明显距离。
我们又做了最后一次尝试，把原博主的生成视频发给它，让它尽量还原。这一版的效果好了不少，整体质感和画面完成度都明显提升。也有几处一眼能看出问题的地方，比如画面中多次出现的光线都像中空的柱子，缺乏真实光照的效果。
这项任务的额度明细共有九条记录，累计消耗6.91%，是五项测试中最高的一项。其中最后一次尝试及随后的调整就消耗了2.88%。
这两个案例放在一起，机械键盘一次完成，只消耗1.46%；《桃花源记》经过多轮修改、两次提供参考素材，累计消耗6.91%，是前者的将近五倍。
回到产品本身，整体来看，五项任务合计消耗约25.48%的月度额度。对于需要反复修改、调用工具才能完成的复杂创作任务来说，这样的消耗已经不算低。
其中《桃花源记》消耗6.91%，苏式老院6.21%，鹈鹕视频5.98%，美术史网页4.92%，机械键盘1.46%。其中，《桃花源记》和苏式老院经过多轮尝试后最终效果仍有改进空间。
这组记录呈现出三类成本：反复修改带来的成本、工具调用带来的成本，以及参考素材与反复生成相关的额外消耗。
产品体验之外，近期kimi的海外宣发争议也不小。
Kimi找到一位海外博主做推广，却要求藏起广告标识，假装是博主自己主动分享的真实体验。
有人评论到：“kimi K3.1 Pro本身是一款很不错的模型，根本不需要招人刷虚假好评，可他们偏要投机取巧耍手段，让整场产品发布观感一塌糊涂。”
按照他的说法，参与活动的创作者需要提前提交文案、获得审核，并被要求不要添加“Paid Partnership”付费合作标签；Brief中还出现了将Kimi与Claude Opus 5.5进行比较的宣传话术。
不过目前相关文件尚未得到核实，无法确认文件由月之暗面还是第三方机构发出。
还有多位海外博主反映，自己发布与Kimi K3.1 Pro有关的内容后，遭到了X平台的处罚。
测评博主Lumina称，自己只是在模型发布前通过Arena测试了模型，既没有收取报酬，也没有获得Kimi官方的提前体验资格，却有两条帖子受到处罚。
据这些博主描述，部分账号需要先添加标签，才能恢复账号或提出申诉。有人为了拿回账号，只能先按要求添加标签，随后删除相关帖子，以免粉丝误以为自己收取了推广费用。
X平台的规则要求，属于付费合作的自然内容必须显示“Paid Partnership”标识，平台尚未公开解释这些账号为何被集中处理。
对外界来说，这让本就存在争议的产品发布增添了新的疑团，在付费推广、独立测评和平台审核之间，究竟发生了什么？
今年6月，前米哈游国际化总裁金雯怡加入月之暗面，负责Kimi出海和全球化，未来将常驻海外。据报道，金雯怡于5月底离开米哈游，完成交接后加入月之暗面。
在米哈游，金雯怡曾长期负责海外发行和国际化业务，参与搭建HoYoverse全球品牌，也经历了《原神》《崩坏：星穹铁道》等产品从中国走向全球的过程。
她加入Kimi，说明月之暗面要解决的问题已经变了：Kimi不再只需要做出更强的模型，还要把模型交付给不同国家的用户，并建立一套能够持续运转的全球化业务。
月之暗面今年年初曾在与投资人的沟通中提到，Kimi海外收入已经超过国内；K2.5发布后，全球付费用户增长了4倍。
Kimi的海外增长来自优秀的模型性能，但一家中国大模型企业的全球化，必然也会伴随着处理本地化、渠道、品牌和信任问题。
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。
♪ We've got a lot of ideas ♪
In the early hours of 24 October, Kimipublish's new model K3.1 Pro. It was officially assigned two major upgrades: the capacity to enter the flagship model ' s first ladder, and the API price was about one third lower than the previous generation.
However, after model publish, Kimi was missing because Token consumption rates were in dispute.
In terms of pricing, officially published API prices have indeed declined.
K3.1 The input price of Pro fell from $20 per million Token to $13 and the output price from $100 to $68. It can be found with C-end users and overseas developers that the actual feeling is not synchronized to cheaper.
The classic “pieron bike” mission, which is tested by the user, is produced for an hour and 47 minutes, 5 hours quota consumes 87 per cent and week quota 16 per cent. And the developers recorded that sending a "hello," and Token counted up to 42,000.
On 26 October, Kimipublish made an apology and a statement of compensation acknowledging that “the problem is with us and should have been repaired earlier”. Two measures were proposed: replacement of and compensation for quota for subscription members during the validity period; and restructuring of the model to reduce unnecessary subAgent calls and output length.
The statement acknowledged that the call mechanisms behind the model were not well controlled.
In the past, the user ' s understanding of model prices was relatively simple: how much token was exported, calculated at unit prices.
But now, a request may also include long context reading, back-office reasoning, mission dismantling, tool call, subAgent collaboration and multiple rounds of self-correction.
The user sends only one sentence, and the model may complete a sequence of actions in the back. As long as there are more Tokens consumed per assignment, even if API prices are reduced, the end cost to the user may increase.
K3.1 Pro, officially known as the “Model of 3 trillion tiers of parameters with the greatest capacity to date”, supports 1 million Tokencontext, located in long-range programming, 3D creative and financial research, and, in these advantage scenarios, easily triggers the type of task that context and MultiAgent call.
The independent research institute Artificial Analysis tested the smart index K3.1 Pro score of 55, ranked fourth in the world, first in the open source model, 11 points higher than K3. According to the results of this list, if the planned open model weights follow, it is also expected to be the highest-scoring open weights model of the list.
These achievements demonstrate that Kimi did indeed complete a capacity leap.
But there is also the real use experience of ordinary users between base model performance and product experience.
In order to observe K3.1 Pro ' s performance in day-to-day tasks, several creative tasks were tested through the Kimi computer application using Pro version subscriptions.
The focus of the test was not on whether a single response was beautiful, but rather on four elements: the first generation of effects, the modification process, the final completion level, and quota consumption.
As of the afternoon of 27 October, the total monthly usage of the Account was 27.03 per cent, of which five test missions combined consumed approximately 25.48 per cent.
Pelican cycling is a classic case that follows almost every new model publish. It's the source of technology, Simon Willison. The title is only one sentence: General an SVG of a pelican riding a vector.
SVG needs to describe the graphics in code, with models that understand the structure of pelicans and bicycles and that address the spatial relationship between wheels, racks, legs and pedals. Why doesn't Simon get along with Pepsi? One is because Simon himself liked it, and the other because the same type of asset was missing on the Internet at the time, and it was more difficult for the model to reproduce the pictures directly from the training data.
As model capabilities progressed, the test became the “opening ceremony” after the new model publish. This time, we have upgraded the title to short video, and we continue to observe, beyond the structure and beauty of the operation, whether it is natural or not, whether the picture is consistent.
We ask the pelicans to smile, to ride their bicycles along the coastline in the sunset, and the waves to shore.
Once the mission was completed, without any additional modifications, the results were good, the pelican expression was lively, and the hair of the body rose with the wind; the waves were natural, and the whole was warm.
quota details show that the K3.1 Pro entry consumes 0.13 per cent and the Plugin entry consumes 5.85 per cent and a total of 5.98 per cent. In other words, the mission did not return to work and approximately 97.8 per cent of the consumption came from the Plugin entry.
Next, we make it a Su-style home.
prompt is “to make a Su-style old courtyard, with a tree of twig, under which there are vines and benches and tables, and a cup of tea”.
The first version of the product is rough, and the whole body feels close to the granule, and leaves are like paper.
In the second edition, we call for the elimination of the sand-drive and the replacement of natural vision and light view in the yard; the use of low-angle sun to strengthen the summer atmosphere; and the replacement of the bench with a rectangle and the continued refinement of other details.
This version addresses the main issues of perspective, couches, and fanning, and the details still differ. The bamboo forest outside the courtyard was too simplistic, the fan was suspended and the plume was embedded in the roots.
The third edition, we have provided a true reference map for it to recreate. With the reference map, the picture is more informative and the problem is focused on the details. The pillar is prominent and the rocks around the pond are close to foam.
The breakdown of quota shows that for the first time, the old Sioux home generated 1.76 per cent, and that subsequent modifications continued to increase by 4.45 percentage points, with the total mission consumption of 6.21 per cent, second in five tests.
The third task is to create a web page on the history of world fine arts. We ask that the representative person and the painting be sufficiently rich, and that the user, by clicking on the person's head or painting, can view the relevant knowledge solution.
The resulting web page is called “ATHÈNE's World History of Fine Arts, Romantic Gallery”, which creates a romantic atmosphere of warm paper, dark brown and soft light and light. Users can screen cultural tracks, characters or paintings, and can roam by chapter, search for keywords, and the movement of people can awaken.
The first edition, which is closer to a select gallery, contains only 13 characters and 14 works, which are in small numbers and are mixed with Western and Central art. When clicking on the " Next item of the guide ", the right text explanation and the left side painting are not synchronized and the painting is shown approximately three seconds later than the text.
After further optimization, the web page was expanded to 30 characters and 30 works, presented in a dual track of Western and East Asian art, with a clear timeline of the art. Click on any person ' s header, painting or axle to read the relevant background. The end of the cell phone can also be browsing, with a balance between beauty and ease of use.
The model entries on the Art History page consume 4.90 per cent and the Plugin entries 0.02 per cent, or 4.92 per cent.
Next, we recreated two bloggers' tests.
The first is the mechanical keyboard, which is required to produce a high-precision model showing the parts structure of the cap, the axis, the locator, the PCB, the shell, and other parts, while supporting the white model, the X-ray lens, the explosion view and the different directions.
This generation did not undergo multiple re-engineering, the main structure was complete, and multiple views could be switched. The breakdown of quota shows two billing records, 1.09 per cent and 0.37 per cent, respectively, for a total of 1.46 per cent, the lowest of the five missions.
The second task is to restore the Peach Flower Book.
The first version of the image is symbolic, the characters and objects are replaced by abstract elements, and the whole quality is close to the first version of the old Sioux court. We then provided a reference map of the mountain to be remade. The second edition still does not capture the spatial changes and narrative levels of the original text, and the picture is clearly at a distance from the meaning of "Peachflower".
We made a last attempt to send it a video of the original blogger and to restore it as much as possible. The results of this edition have improved considerably, with a marked improvement in the overall quality and picture completion. There are also a few places where the problem can be seen at first sight, such as the fact that the light that has appeared repeatedly in the picture is like a pillar in the middle and lacks the effect of real light.
The quota detail for this task contained nine records, with a cumulative consumption of 6.91 per cent, the highest of five tests. The last attempt and subsequent adjustments consumed 2.88 per cent.
The two cases were taken together, with the mechanical keyboard being completed at one time and only 1.46 per cent consumed; and the Peach Flower Source Book, which was revised on multiple rounds and twice provided reference asset, consumed a cumulative 6.91 per cent, almost five times the former.
Returning to the product itself, the five missions together consume about 25.48 per cent of the monthly quota. Such consumption is no longer low for complex creative tasks that require repeated modifications and the use of tools.
Of these, Peach Flowers consume 6.91 per cent, Su-style 6.21 per cent, Pelican video 5.98 per cent, Art History web page 4.92 per cent and mechanical keyboard 1.46 per cent. Among them, there is still room for improvement in the end effects of the Peach Flower Book and the Su-style Home after several rounds of experimentation.
This group presents three types of costs: the cost of repeated modifications, the cost of tool call and the additional consumption associated with the reference to asset.
In addition to the product experience, the recent overseas dispute over Kimi has been significant.
Kimi found an overseas blogger to promote, but asked to hide the logo, pretending to be the real experience he had voluntarily shared.
It was commented: “Kimi K3.1 Pro is a very good model in itself, and there is no need to draw a false and positive opinion, but they tend to speculate on the means of playing, and make the whole product publish look a mess”.
According to him, the creators of the event were required to submit their papers in advance, to obtain an audit and to be asked not to add the “Paid Partners” paid cooperation label; there was also a propaganda campaign in Brief to compare Kimi with Claude Opus 5.5.
However, the relevant documents have not yet been verified and it is not possible to confirm whether they were issued by the dark side of the month or by a third party.
A number of overseas bloggers also reported that their own publish content related to Kimi K3.1 Pro had been punished by Platform X.
Lumina, the critic, stated that she had tested the model through Arena only before the model publish, which was neither remunerated nor qualified for the official Kimi experience, but that two posts had been punished.
According to the organizers, some accounts would need to be labelled before they could be reinstated or a complaint could be filed. In order to recover the account number, one can only add the label as required and then delete the relevant post so that fans do not mistakenly assume that they have been charged for the promotion.
The rules of Platform X require that the natural content of paid cooperation be shown as the “Paid Partnership” logo, and the Platform has not publicly explained why these accounts have been centralized.
To the outside world, this added a new mystery to the disputed product publish, and what happened between paid promotion, independent evaluation and platform review?
In June of this year, Kimi was added to the dark side of the moon by Kimi, former President of Miha Internationalization, and will be based abroad in the future. She reportedly left Miha at the end of May to join the dark side of the month after the handover.
In Miha, Kim had long been in charge of overseas distribution and internationalization, had been involved in the construction of the HoYoverse global brand, and had gone through the journey from China to the world of products such as The Gods of the Nativity: Star Dome.
She joined Kimi to show that the issues to be addressed in the dark side of the moon have changed: Kimi no longer needed to make more robust models, but also to deliver them to users in different countries and to create a global business that would be sustainable.
The dark side of the month mentioned in a communication with investors early this year that Kimi had more overseas than domestic revenue; since K2.5publish, the number of paid users worldwide has increased fourfold.
Kimi ' s overseas growth comes from good model performance, but the globalization of a Chinese LLM enterprise is bound to be accompanied by localization, access, branding and trust.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.