文 | 字母AI
K3.1 Pro发布后不到48小时，月之暗面就放出公告称“模型存在的限速、额度无故虚耗问题，我们深表歉意。这些问题全是我们的责任，本该更早推动解决。10月26日起，所有档位订阅用户返还月度总额的25%。我们还调整了产品运行参数，砍掉不必要的子Agent调用、压缩冗余输出长度，让额度更耐用。”
在这次改动后，最明显的体验就是，K3.1 Pro没有刚发布那会“烧额度”了。
回到模型发布当天，整个AI圈子都倒吸一口凉气，性能超越OpenAI所有模型，仅落后Opus 5.5半个脚掌。第三方评测机构ModelCompare发文称，中美AI差距已从3、4个月缩减为仅剩31天。
然而这场狂欢还没撑过两天，K3.1 Pro的评价就从“特别好评”，变成了“多半好评”。杨植麟也被冠上了“神鬼二相性”的外号，神在于模型性能突破天际，鬼在于K3.1 Pro用起来多少有点“肉疼”。
反馈最多的就是额度问题。社区多位用户称，几乎没怎么用，就已经花光了K3.1 Pro的5小时额度。就拿我自己来说，199块钱的档位，两天一共生成不到10个3D模型和几个时长不超过30秒的演示动画，月额度已经跑完了一半。
哪怕是一些简单的任务，K3.1 Pro也会调用大量的子Agent来思考、推理，其输出结果并没有提升很大，却因为这些子Agent，导致用户浪费了很多额度。
杨植麟也挺无奈，提升模型的智能水平必然需要更多的token。
这也就导致K3.1 Pro虽然比K3便宜了30%，却因为额度给得太少，导致最后并没有多少人真正感受到了“便宜”。
就在K3.1 Pro发布的前一周，DeepSeek发布了V4.1 Flash-1017，它的思路完全相反：不拼上限，而是把性能、速度、输出长度一起优化，价格保持不变。
V4.1 Flash-1017完美体现了梁文锋的产品哲学，自发布至今仍然是“好评如潮”。
或许杨植麟就该在产品上，该跟梁文锋取取经。
额度是月之暗面最大的伤
K3.1 Pro发布后，社区里有很多人都在讨论额度的问题。比如有这样一条高赞评论，发帖人称“它大概只用了10分钟，5小时的额度窗口就已经满了”。
Linux.do上，有一位199套餐用户，他表示只发了一句让模型审查项目架构的指令，Kimi拉起5个子 Agent，十几分钟后5小时额度就满了。
微博用户@这很盒里吧发帖表示，在WorkBuddy里使用K3.1 Pro，一个简单任务还没跑完，就花掉了超过450的积分。
而实际上，WorkBuddy的99块钱标准版，一个月总共也就4000积分。相当于一个简单任务，跑了月额度的10分之1。@这很盒里吧还表示，K3.1 Pro目前消耗积分的倍率是0.97，也就是说，实际上积分消耗要更少一些。
Artificial Analysis认为，之所以K3.1 Pro如此消耗额度，关键就在于token效率上。
在测试中发现，K3.1 Pro跑完整套Intelligence Index测试集，一共要输出约1.5亿个token，同价位模型的中位数是8700万。
同一套测试，GPT-6.1 Sol只用了约6700万，GPT-6 Astra约6000万。按OrcaRouter的换算，同样的任务K3.1 Pro的输出量大约是Astra的2.5倍。
Artificial Analysis还表示，K3.1 Pro的速度也低于同类的中位数，大约每秒70多到80个token。另外，K3.1 Pro只支持思考模式，关闭思考的请求会被直接拒绝，每一次调用都在为推理付费。
虽然K3.1 Pro推出了高速（Highspeed）版本，但是月之暗面官方表示，高速版只加快模型输出。任务里工具调用（读写文件、执行命令等）与脚本执行的耗时不受影响，如果这部分占比大，整体提速就不明显。
虽然K3.1 Pro相较于K3便宜了30%，可Artificial Analysis在测试中发现，K3.1 Pro的单次任务成本并没有便宜到哪里去。
K3.1 Pro平均单次任务的花费为2.11美元，GPT-6 Astra为3.26美元。而K3.1 Pro的token价格只有Astra的五分之一，可落到实际花费上，只省了大约35%。
这一切可能和K3.1 Pro的训练有关。
杨植麟认为，要想让模型实现更长的输出，那就必须得在整个过程中始终保持严谨，反复检查、修改、自我修正。从K3开始就是如此，当时的官方技术文档里写到，K3会在输出中表现出过度谨慎、反复自我审视。
Fireworks AI官方研究博客《Introducing Ember-1》中提到，就算拿掉了K3反复检查的过程，对模型性能几乎没有影响。换句话说，K3有点太严谨了。
到了K3.1 Pro，模型不仅要对任务严谨，还对整个环境严谨。
K3.1 Pro官方技术报告里提到过一个案例，模型模拟人学习行走的过程，K3.1 Pro发现手臂动作始终不理想、调整奖励也没有解决，于是回头检查模拟器，修掉了36处肩部肌肉力臂的计算错误，再重新训练，最终才让效果更好。
这也是为什么K3.1 Pro虽然性能很高，然而输出更长，token更多，因为它要对几乎一切保持严谨，甚至连环境都是如此。
杨植麟烧token，梁文锋则是省token。
DeepSeek称V4.1 Flash 1017的迭代为“帕累托改进”。这是一个经济学概念，它指的是一件事改完之后，至少有一个方面变好了，同时没有任何一个方面变差。
放在模型里体现出的，就是模型的各项基准分数更高，任务完成更快，输出更短，价格也没有变。
DEV社区里就有人做过测试，说DeepSeek V4.1 Flash 1017相比较4.1 Flash，输出变短、任务变快，每完成一个任务的成本，反而下降了。
其实V4.1 Flash本身是个极其“话痨”的模型，OrcaRouter上就有实测表示，同样是在Artificial Analysis那套Intelligence Index上，V4.1 Flash要输出2.5亿token，比K3的1.6亿还多56%。
但是由于V4.1 Flash的Causal Encoder-Decoder架构，552B总参数但prefill只激活8B、decode只激活16B。KV缓存被压到每token只有890字节，显存占用降到上代的四分之一、SSD占用上代的八分之一。
所以DeepSeek跑分更强、任务完成更快、输出还变短，梁文锋不愿意牺牲掉某一项属性换来另一项。
Agent是服务AGI的一环
梁文锋曾经说过，“现在投资人看到的最多的是Agent，但对于我们这些研究的⼈来讲，现在看到更多的是学习，以及怎么解决学习这个问题。”
言外之意，梁文锋觉得Agent只是AGI的一环而非全部。
2025年的时候，杨植麟登上了张小珺《商业访谈录》。对于纯推理模型和Agent，杨植麟提出了一个缸中之脑理论：“想象一个鱼缸，你把一个脑子放在里面，跟外界没有联系。它只是在自己大脑里面想，一直想，不需要跟外界产生任何交互，就能解一道题。”
对于Agent，杨植麟说：“它不再是‘缸中之脑’，是跟外界有交互的。我的下一步行为，是根据交互中得到的反馈，和外界给我的新状态的更新有关系。”
缸中之脑就代表了o1那样的强推理模型，可杨植麟想要的，是让模型能够使用更多的工具，进行更多轮的交互。
不过无论梁文锋和杨植麟怎么说，那都是他们在产品哲学方面对Agent的定义，真要看模型Agent如何，还得放到真实任务里。
除了Artificial Analysis，社区也经常用“致知｜大模型智力长期追踪榜单”（猫榜）来测试模型的水平。
它把模型的能力拆成推理、Agentic和视觉三个板块。在Agentic这一页，用的是一组需要长流程、多约束和工具协作的任务。左边的数字和字母，表示扣分数和档位，页面上写着“1/A 代表扣分数/档位”，数字越小越好；右边的￥金额，看起来是模型跑完这道题的花费。
相比上一代K3，K3.1 Pro的进步非常明显。K3在Rust App上是“Failed”，也就是不能实现，K3.1 Pro则拿到了23分、C档；K3在Science上是26分、C+，K3.1 Pro提升到了16分、B档；Animation也从22分提到了17分。
这个结果，和月之暗面官方强调的“Agent执行能力大幅提升”是吻合的。
然而在这个榜单上，K3.1 Pro却输给了DeepSeek V4.1 Flash 1017。
Rust App上，DeepSeek V4.1是12分、B档；K3.1 Pro是23分、C档；Simple Model这一项，K3.1 Pro是5分，DeepSeek是10分，两者同为B+档，Kimi略好一些；Animation这一项，双方都是17分、B档，打平；Science 一项，DeepSeek是13分，Kimi 是16分，同为B档，DeepSeek略好。
另外，K3.1 Pro在Data Process和Metal两项还是“Pending”，意味着完整的结论目前还不能下。
明明在Artificial Analysis上直逼Opus 5.5的K3.1 Pro，却在这里评分低于DeepSeek？
这是由于猫榜有个规则，“模型不限制Max Token，但根据题目难度对超长思考进行扣分惩罚。”等于你花的token越多，就越给你扣分，K3.1 Pro吃了大亏。
在Rust App上，DeepSeek花了12.32元，K3.1 Pro花了221.21元。Simple Model上，Kimi花了12.60元，DeepSeek是5.88元。Animation上，Kimi是38.77元，DeepSeek是10.77元。Science上，Kimi是25.80元，DeepSeek是6.21元。
还有一点，K3.1 Pro的技术博客里写到，模型会“过度主动（excessive proactiveness）”，遇到模糊意图会替你做决定、过度即兴。尤其是对于Agent来说，K3.1 Pro更愿意去根据提示词来自由发挥。
可猫榜里面，像是Rust App这类任务，它的环境反馈链极长、需求边界完全卡死，会严格限制模型的发挥。
梁文锋在做的，是让每一次Agent调用都稳定、便宜；杨植麟在做的，是让最难的那几个场景，看起来足够惊艳。
月之暗面需要“克制”
K3.1 Pro的3D能力是发布的重点之一。你可以给模型几张图片，不需要交代任何事情，它就能将其转换成3D建模。
发布后的第二天，社区上出现了一个高赞帖子。发帖人称，他发现Kimi内置了一个名为model3d_generation的技能，会运行一段Python脚本，通过外部API的服务来生成3D模型。
发帖人特意禁用了这个技能，让K3.1 Pro只用自己的代码生成，质量明显差了一截。更关键的是，脚本里支持三个模型版本：v3.1、v3.0和v2.5，这和3D生成公司Tripo AI的模型版本完全吻合。
事实上，早在K3.1 Pro发布之前，Kimi Work桌面端就已经把Tripo 3D纳入了插件生态。
月之暗面资源中心10月17日更新的《如何制作具有逼真效果的3D图像设计》一文中明确写道，Kimi Work“支持 Blender、Tripo 3D及其他插件”。
官方10月14日的博客《10款值得尝试的免费和付费AI 3D图像生成器》中也提到，Kimi Work已经集成了 Tripo 3D。
其实月之暗面自己发布的技术博客里明确表示K3.1 Pro在游戏引擎里工作时，会“结合网页搜索和3D生成”来创造完整的体验。BenchCAD的92.7%，标注的也是“借助工具”。
这意味着，官方并没有说这些成果是模型“徒手”完成的。
请“外援”这件事，并没有什么错。调用工具，原本就是Agent能力的一部分，一个会找到合适工具并把它接进工作流的模型，比一个什么都自己硬算的模型更有用。
但问题就在于，杨植麟想要的太多。除了3D建模和游戏引擎，K3.1 Pro所覆盖的还有代码、长程Agent、金融研究等等。
每一个方向单独看，都有道理，也都有真实的进展。
Artificial Analysis的数据显示，K3.1 Pro在科学推理上是34.3%，同期 Opus 5.5是59.0%，GPT-6 Astra是63.3%。在不用工具的HLE测试里，K3.1 Pro是51.1%，低于 Opus 5.5的59.8%；长上下文推理，也从K3的88.7%略降到了86.3%。
战线拉得太长，每一条线上能分到的资源就变少了。
相较之下，梁文锋就非常克制，他曾经说过：“这也是我们克制的一部分。我希望这些商业机会能够让其他⼈来做，我们希望这些商业机会、怎么用这个AI ，是整个社会、我们所有的合作伙伴⼀起来做，大家⼀起来分享这个收益，而不是我想独吞，这不可能。”
DeepSeek并不是什么都不做，它同样有搜索和抓取API，有各种产品化的动作。区别在于，每一件事都被放在“服务AGI主线”这个前提之下，不服务主线的事，哪怕是好生意，也宁愿不做。
克制不是佛系，它本质是一套资源分配的方法。先决定什么不做，再决定怎么做好。
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。
Letter AI
Less than 48 hours after K3.1 Propublish, the dark side of the month released the announcement that “the speed limit of the model's existence and quota's unnecessary exhaustion are problems, and we apologize. Those issues are all our responsibilities and should have moved forward earlier. Effective 26 October, all slot subscribers returned 25 per cent of the monthly total. We also adjusted the parameters for the running of the product, cutting off the unnecessary subAgent call, compressing the redundancies output length to make quota more durable."
The most obvious experience after this change is that K3.1 Pro did not have just publish that would “burn quota”.
On the day of the return to the model publish, the whole of the AI circle was pumped into the cool air, capable of surpassing all Open AI models, and was only 5.5-and-a-half feet behind. According to a third-party evaluation agency, Modelcompare, the gap between China and the United States of America has been reduced from 3 to 4 months to 31 days.
However, before the party lasted two days, K3.1 Pro's evaluation went from “specially good” to “mostly good”. Yang Shizuku also has the nickname "The Two of the Ghosts", where God's model works beyond the sky, and K3.1 Pro uses a bit of "the flesh pain."
The biggest feedback was on quota. A number of community users stated that they had spent 5 hours of quota on K3.1 Pro, almost without much use. In my case, the 199-dollar slot produced less than 10 3D models and several demonstration animations of less than 30 seconds in two days, and the month quota is halfway through.
Even for simple tasks, K3.1 Pro would call on a large number of sub-Agents to think, reason, whose output did not go up much, but because of these sub-Agents, users wasted a lot of quota.
It's not easy for Yang to raise the level of intelligence of the model.
This has led to K3.1 Pro being 30 per cent cheaper than K3, but because quota has been too little, resulting in less than a few people actually feeling “cheap”.
Just one week before K3.1 Propublish, DeepSeekZ94Z took V4.1 Flash-1017, the idea being the opposite: not to spell ceilings, but to optimize performance, speed, output length and price.
V4.1 Flash-1017 represents the perfect product philosophy of Liang Wenbei, which has remained “goodly received” since publish.
Maybe Yang Shizuku should be on the product. He should learn from Liang Wenbing.
quota is the worst wound to the dark side of the moon.
After K3.1 Propublish, there was a lot of discussion in the community about quota. For example, in one of the high-profile comments, the author stated that “it took about 10 minutes and that the five-hour window of quota was full”.
On Linux.do, there was a 199 set of food users who indicated that only one order had been issued to allow the model to review the project structure, and Kimi pulled up five items, Agent, and quota was full in 10 minutes.
Weibo users @ This is a box of cards that says that using K3.1 Pro in WorkBuddy, a simple task is spent more than 450 points before it was finished.
In fact, the 99-dollar standard version of WorkBuddy amounts to 4,000 points a month. It's equivalent to a simple mission, running one tenth of the month quota. @This box also indicates that K3.1 Pro currently consumes 0.97, which means that it actually consumes less.
According to Artificial Analysis, the key to K3.1 Pro ' s consumption of quota is token efficiency.
The test found that K3.1 Pro ran the complete set of Inteligence Index tests, with a total output of about 150 million tokens, with a median of 87 million for the same price model.
In the same set of tests, only about 67 million GPT-61 Sol and about 60 million GPT-6 Astra. According to convert of OrcaRouter, the output of K3.1 Pro for the same task is approximately 2.5 times that of Astra.
Artificial Analysis also states that K3.1 Pro is also below the median of the same type, with about 70 to 80 tokens per second. Furthermore, K3.1 Pro only supports thinking patterns, and requests to close thinking are directly rejected, and every call is made for reasoning paid.
Although K3.1 Pro has launched a high-speed (high-speed) version, the dark side of the month officially indicates that the high-speed version only accelerates model output. tool call (reading and writing documents, execution orders, etc.) and script execution in the mission are not affected by the time-consuming nature of the operation, and if this proportion is larger, the overall speed increase is not apparent.
Although K3.1 Pro was 30 per cent cheaper than K3, Artificial Analysis found that K3.1 Pro ' s single mission costs were not cheap.
K3.1 On average, Pro spent $2.11 on individual missions and $3.26 on GPT-6 Astra. The price of K3.1 Pro token, which is only one fifth of Astra ' s price, falls to real costs, saving only about 35 per cent.
This could all be about K3.1 Pro training.
Yang believed that if the model was to produce a longer output, it would have to be carefully checked, modified and self-corrected throughout the process. This has been the case since K3, when it was written in the official technical files that K3 showed excessive caution and self-examination in the output.
The official Fireworks AI research blog, Introducing Ember-1, mentions that even if the K3 process is removed, it has little impact on model performance. In other words, K3 is a bit too strict.
At K3.1 Pro, the model is not only mission-specific, but environment-wide.
K3.1 An official technical report of Pro mentioned a case where model simulators learned to walk, and K3.1 Pro found that arm moves were always not ideal, that adjustment incentives were not resolved, and then looked back at the simulators, which had fixed the computational errors of 36 shoulder muscle arms and retrained to make the effects better.
This is also why K3.1 Pro, although highly performing, exports are much longer and token more, because it has to be strict about almost everything, even the environment.
Yang Shih burned token, and Liang Wenbei Province.
DeepSeek describes V4.1 Flash 1017 as an iterative “Pareto Improvement”. This is a concept of economics, which means that when one thing is changed, at least one aspect is better, and none is worse.
What is reflected in the model is that the baseline scores of the model are higher, tasks are completed faster, outputs are shorter and prices remain unchanged.
Tests were carried out in the DEV community, stating that DeepSeek V4.1 Flash 1017 compared 4.1 Flash, with shorter output, faster tasks and lower costs per mission.
In fact, V4.1 Flash is itself a very “talking” model, as measured on OrcaRouter, and also on Artificial Analysis's Intelligence Index, V4.1 Flash is exporting 250 million tokens, 56 per cent more than K3 billion.
However, as a result of the Causasal Encoder-Decoder structure of V4.1 Flash, the overall parameters of 552B are only activated by 8B, but only 16B is activated by prefil. KVcache was crushed to 890byte per token, VRAM to one quarter of the previous generation and SSD to one eighth.
As a result, DeepSeek has a stronger run, faster mission completion and shorter output, and Liang Wenbing is reluctant to sacrifice one attribute for another.
Agent's part of the AGI service.
As Liang Wenfeng once said, “Agent is now the most seen by investors, but for those of us who study, more is now seen in learning and how to solve it”.
Beyond that, Liang Wenfeng feels that Agent is only part of AGI, not all.
In 2025, Yang Shih-hyun posted a commercial interview. As for pure reasoning models and Agent, Yang Shizuku suggested a brain theory in a tank: "Imagine a fish tank, you put a brain in it, you have no contact with the outside world. It's just thinking in its own brain, always thinking that it doesn't have to interact with the outside world to solve a problem."
In the case of Agent, Yang Shih said, “It is no longer the brain in the tank, it interacts with the outside world. My next step behavior is based on the feedback received during the interaction and is related to the new state of affairs that the outside world has given me.”
The brain in the tank represents a strong reasoning model like O1, but what Yang wants is to give the model more tools and more rotational interactions.
But whatever Liang and Yang Shizuku say, that's their definition of Agent in terms of product philosophy. It depends on the model Agent, which has to be placed in real task.
In addition to Artificial Analysis, communities often test the level of models by using the “Today the LLM Intellectual Long-Term Tracking List”.
It breaks down the capabilities of the model into three sections: reasoning, agency and vision. On the Agenic page, a set of tasks required long pipeline, multi-binding and tool collaboration was used. The number and letter on the left, which represent the number of points and slots, is written on the page “1/A for the number of points/slotting”, the smaller the number, the better; the amount on the right appears to be the cost of the model running through the subject.
The progress of K3.1 Pro is clear compared to the previous generation of K3. K3 was “Failed” on Rust App, i.e. not achievable, while K3.1 Pro received 23 points and C; K3 was 26 on Science, C+, K3.1 Pro was promoted to 16 and B; and Animation mentioned 17 points from 22.
This result coincides with the official emphasis on “a significant increase in the capacity of Agent to implement”.
On this list, however, K3.1 Pro lost to DeepSeek V4.1 Flash 1017.
According to Rust App, DeepSeek V4.1 is 12, B; K3.1 Pro is 23, C; Simple Model is 5; DeepSeek is 10, the same as B+ and Kimi is slightly better; Animation is 17, B is even; Science is 13, DeepSeek is 16, same as B, DeepSeek is slightly better.
Furthermore, K3.1 Pro's “Pending” or “Data Proces” or “Pending” means that the complete conclusion is not ready at this time.
The K3.1 Pro of Opus 5.5 on Artificial Analysis is below DeepSeek?
It's because the cat's list has a rule, "The model doesn't limit Max Token, but the deduction of a long thought according to the difficulty of the subject." The more token you spend, the more points you get, the more K3.1 Pro loses.
On Rust App, DeepSeek spent $12.32 and K3.1 Pro $221.21. On Simple Model, Kimi spent $12.60 and DeepSeek $5.88. On Animation, Kimi was $38.77 and DeepSeek 10.77. On Science, Kimi is $25.80 and DeepSeek is $6.21.
Another point is that K3.1 Pro’s technical blog says that models are “excessive initiatives” and that vague intentions make decisions and improvise for you. Especially for Agent, K3.1 Pro prefers to play freely under prompt.
But on the cat list, like Rust App, it has an extremely long environmental feedback chain and a completely stuck demand boundary, which severely limits the use of models.
What Liang Wensel is doing is making every call by Agent stable and cheap; what Yang is doing is making the hardest scenes look amazing enough.
The dark side of the moon requires restraint.
K3.1 Pro ' s 3D capability is one of the priorities of publish. You can give the model a few pictures, and it doesn't have to say anything, and it turns it into a 3D model.
The day after publish, a high-profile post appeared in the community. The sender stated that he had discovered that Kimi had built a skill called model3d generation and would run a Python script to generate a 3D model through external API services.
The sender specifically banned the technique, allowing K3.1 Pro to be generated only by its own code, with a marked difference in quality. More critically, three model versions are supported in the script: v3.1, v3.0 and v.2.5, which is fully consistent with the model version of Tripo AI, a 3D generator.
In fact, before K3.1 Propublish, Tripo 3D had been included in the plugin ecology at the end of Kimi Work.
Kimi Work “Support for Blender, Tripo 3D and other plugins” was clearly stated in the 17 October update of the Centre for the Dark Side of the Moon, “How to Make 3D Images with Real Effects”.
The official blog of October 14th, Free and paid AI 3D Image Generators, also mentioned that Kimi Work has assembled Tripo 3D.
In fact, the dark side of the moon itself, in its technical blog publish, clearly states that when K3.1 Pro works in the game engine, he creates a complete experience by "comprising web search and 3D generation." 92.7 per cent of BenchCAD, also labelled as “tool-based”.
This means that the official did not say that the results were done by “hand”.
There is nothing wrong with asking for “outside assistance”. The call tool, originally part of Agent's ability, is a model that will find the right tool and feed it into workflow, which is more useful than a model that does everything by itself.
But the problem is that Yang Shih wants too much. In addition to 3D modelling and game engines, K3.1 Pro covers code, long-range Agent, financial research, etc.
Each direction alone makes sense and there is real progress.
Data from Artificial Analysis show that K3.1 Pro was 34.3 per cent in scientific reasoning, while Opus 5.5 was 59.0 per cent and GPT-6 Astra was 63.3 per cent during the same period. In HLE tests without tools, K3.1 Pro was 51.1 per cent, down from 59.8 per cent in Opus 5.5; the long context reasoning also slightly decreased from 88.7 per cent to 86.3 per cent in K3.
The battle lines are too long, and the resources available on each line are diminishing.
By contrast, Liang Wenfeng was very restrained, as he once said: “It is also part of our restraint. I want these business opportunities to be done by others, and we want these business opportunities, how to use this AI, by society as a whole, by all of our partners, to share the benefits, not by me alone, which is impossible.”
DeepSeek does not do anything, it also does search and capture API, and it does all kinds of product-oriented actions. The difference is that everything is placed under the premise of “serving the AGI main line”, and that what does not serve the main line, even if it is a good business, prefers not to do it.
Restraint is not a Buddhist system; it is essentially a way of allocating resources. Let's decide what to do, then what to do.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.