Kimi K3.1 Pro额度争议背后：模型越强，算力越紧

[[**Kimi K3.1 Pro的性能已接近全球顶尖水平，却因订阅额度争议暴露出算力瓶颈：模型越强，Agent越耗Token，中国大模型公司的商业化与普及仍受算力约束。**

**要点**

K3.1 Pro在Terminal-Bench 4.0上得分65.7%，接近Claude Opus 5.5；但199元和699元套餐的用户抱怨额度不足，开发者测算旧套餐的API价值约为新套餐的1.94倍。

**2. API降价，Agent仍把算力需求推高**

K3.1 Pro每百万Token的输入/输出价格降至13元和68元，低于K3的20元和100元；但复杂Agent可消耗数百万Token。2026年6月国内日均Token约500万亿，较2025年12月增长近4倍。

**3. 算力短缺长期约束独立模型公司，国产供给仍难完全解决**

本文来自微信公众号：01Founder，编辑：Max，作者：一直在路上的Max

10月24日凌晨，月之暗面发布了新一代模型Kimi K3.1 Pro。

这是一个拥有3万亿级参数、支持100万Token上下文的大模型。

在官方公布的Terminal-Bench 4.0测试中，它取得了65.7%的成绩，已经非常接近Claude Opus 5.5这样的全球顶尖模型。

对于一家成立仅仅几年的中国创业公司来说，这样的成绩已经足够惊人。

两年前，中国的大模型公司还在苦苦追赶GPT-4，如今，它们已经能够在全球最前沿的技术舞台上，与那些拥有庞大算力和资金储备的美国公司正面竞争。

然而，这场本应迎来掌声的发布，却出现了一幕颇为讽刺的景象。

模型越来越聪明，用户却发现自己越来越用不起了。

在K3.1 Pro发布之后，社交媒体上很快出现了不少关于会员额度的抱怨。

一些用户购买了每月199元的Pro套餐，准备体验新模型的编程和Agent能力，结果没运行几个复杂任务，就发现自己已经接近使用上限。

与此同时，关于Kimi在新模型发布前后调整会员额度的讨论也开始出现。

一位开发者使用两个699元的订阅账号，通过49次API请求进行测试，按照相同的Token结构进行折算，发现旧套餐能够获得的API价值约为新套餐的1.94倍。

虽然这一结果不能直接代表所有用户的实际额度变化，但仍然引起了不少讨论。

更有意思的是，Kimi这次发布的新模型，API价格其实还下降了。

根据官方定价，上一代K3每百万Token的输入和输出价格分别为20元和100元，K3.1 Pro则降到了13元和68元。

单看API价格，Kimi依然在努力维持中国模型一贯的成本优势。

可对于购买订阅套餐的用户来说，模型能力提升了，单位Token价格降低了，真正使用时却需要更加谨慎地计算额度。

模型公司拿出了一张漂亮的成绩单，用户却开始抱怨自己的账单。

看到这些讨论的时候，我突然想起了2024年的一次饭局。

2024年的Kimi，大概是中国最风光的大模型创业公司。

那时候，ChatGPT已经向全世界展示了大语言模型的潜力，但国内真正能够让普通用户记住名字、主动使用的AI产品还没有几个。

Kimi凭借长文本能力迅速出圈，一度成为中国大模型行业最受关注的产品之一。

它的用户增长、产品体验，以及背后月之暗面的融资故事，都让这家公司站在了聚光灯下。

对于那个时期的创业者而言，Kimi几乎代表着中国大模型商业化的一种可能。

恰好当时我有一位朋友正在做企业服务，手里有一些B端客户，希望能够接入Kimi的模型能力，为企业提供定制化的AI应用。

于是，他通过熟人介绍，约了月之暗面的一位联合创始人吃饭，希望能够聊聊合作的可能性。

因为他对模型训练和技术侧的情况不太了解，便特地喊我作陪。

在见面之前，我觉得这原本应该是一件顺理成章的事情。

毕竟那时候，整个大模型行业都在讨论商业化，如何找到真正愿意付费的客户，如何让模型走出实验室，成为企业生产流程的一部分，几乎是所有创业公司绕不开的问题。

而我们眼前恰好就有这样的客户，他们拥有真实的业务需求，也愿意为模型服务支付费用。

一家站在行业风口上的创业公司，手里握着当时国内最受欢迎的大模型产品之一，外面有大量资本希望投资它，也有企业愿意购买它的服务，可当真正的商业机会摆在面前时，它却因为算力不足而无法承接。

当然，我能够理解对方当时的选择。

2024年的Kimi正在快速增长，有限的计算资源需要优先保证自有产品的稳定运行。

对于一家仍在技术和商业化之间摸索的创业公司而言，如何分配这些资源，本来就是一道困难的选择题。

那顿饭之后，我们没有继续推进合作，这件事也逐渐淡出了我的记忆。

直到两年后的今天，Kimi已经从当年主打长文本能力的聊天助手，成长为能够执行复杂编程和长程Agent任务的前沿模型，月之暗面的技术实力和模型规模，也早已不可同日而语。

可是，当我看到用户因为额度不足而抱怨K3.1 Pro的时候，当年那句“算力不够”又一次浮现在我的脑海里。

两年前，有企业客户愿意付费使用Kimi，月之暗面却因为算力紧张，最终没有承接这笔生意。

两年后，Kimi已经训练出了世界一流的大模型，普通用户也愿意每个月花199元甚至699元购买服务，却发现自己仍然需要小心翼翼地分配每一次调用。

今天的会员额度争议涉及具体的产品定价、推理成本和容量管理，我们还无法仅凭外部的使用体验，判断Kimi内部究竟存在多大的算力压力。

但这两件相隔两年的事情，却让我开始思考一个更大的问题。

中国的大模型公司已经有能力造出世界一流的模型，可我们究竟还需要多少算力，才能让这些模型被世界充分使用？

理解这个问题，需要先看看过去两年，大模型究竟发生了多大的变化。

2024年，我们与大模型的关系还很简单。

大多数时候，我们坐在电脑前，向ChatGPT或者Kimi提出一个问题，等待它给出答案。

无论是翻译文章、整理资料，还是编写一段代码，模型的工作通常都伴随着一次对话的结束而告一段落。

但到了2026年，大模型正在从两个方向发生变化。

首先是模型本身越来越庞大。

如果回到两年前，100B、120B这样的千亿参数规模，已经足够让一家模型公司大书特书。

可到了今天，千亿参数早已算不上什么稀奇事，头部模型的规模正在向万亿级迈进。

Kimi K3的总参数量达到2.8T，相当于28000亿；

智谱GLM-5.3达到753B，也就是7530亿；

DeepSeek V4 Pro则进一步达到1.6T，相当于16000亿。

当然，这些模型普遍采用了MoE架构，每次推理只需要激活部分参数，单次计算量远小于总参数规模所暗示的数字。

但对于负责部署这些模型的公司而言，更庞大的模型权重、更复杂的并行计算，以及动辄百万Token的上下文窗口，仍然意味着更高的显存、互联带宽和基础设施要求。

而模型规模的扩张，还只是故事的一半。

其次是Agent的普及正在彻底改变人们使用大模型的方式。

如今的用户已经开始习惯把完整的工作交给模型。

一个Coding Agent接到任务后，可能需要阅读几百个源代码文件，理解整个项目的结构，然后修改代码、调用工具、运行测试，再根据报错信息反复调整自己的方案。

如果任务足够复杂，它还可能启动多个Subagent，让不同的模型实例同时处理不同的问题。

从用户的角度来看，整个过程或许只有一句简单的指令，但在这句话背后，模型可能已经连续工作了几十分钟，进行了上百次推理和工具调用，消耗了数百万甚至更多Token。

于是，一个更加棘手的问题出现了。

模型本身正在变得越来越庞大，而我们要求它完成的工作也在变得越来越复杂。

每一次任务可能消耗更多Token，每一个Token背后又需要越来越精密的计算基础设施，两股力量叠加在一起，正在把整个行业的算力需求推向新的高度。

这让我想到电力刚刚普及的年代。

最初，人们只是希望用电灯替换煤油灯，获得一种更加便宜、方便的照明方式。

但随着电力技术的发展，电动机开始进入工厂，电梯、轨道交通和各种电器相继出现，人们对电力的需求也随之不断扩大。

一盏电灯消耗的电力越来越少，但人类已经开始用电驱动整座城市。

技术进步降低了单次使用的成本，却也打开了一个远比过去庞大得多的需求市场。

今天的大模型正在经历类似的变化。

过去我们只是偶尔向模型提出一个问题，如今却希望它能够像真正的员工一样，连续工作几个小时，甚至同时运行几十个Agent。

模型每变聪明一点，我们就会发现更多值得交给它完成的工作，而这些新工作，又需要更多算力来支撑。

10月23日，中国信通院发布了一份《大模型词元服务性能监测报告（2026年）》，其中的两组数据，恰好为这种变化提供了注脚。

In June 2026, domestic average daily Token consumption had already reached about 500 trillion, nearly 4 times that of December 2025.

At the same time, the normalized average price of domestic large model services rose from 6.8 yuan in the second half of 2025 to 13.46 yuan in 2026, an increase of nearly 100%.

The normalized average price here is a statistical result after uniformly processing different models and billing methods. It reflects changes in the market price structure and cannot be directly understood as all models' APIs having doubled in price.

But looking at these two sets of data together, the changes taking place across the industry are already quite obvious.

Over the past two years, one of the most important stories in China's large model industry has been the price war.

From several leading vendors repeatedly cutting API prices to DeepSeek disrupting the global market with highly competitive pricing, Chinese models gradually established a distinct industry impression: capabilities are getting stronger, and prices are getting cheaper.

However, as models gradually transform from chat tools into production tools capable of sustained work, the economic logic of this competition has also begun to change.

In the past, people mainly focused on how much it cost for a model to generate one Token.

Today, as model scale continues to expand and Agent tasks become increasingly complex, how much computing resources are needed to complete a piece of work has begun to become a more important question.

This also explains why, even though Kimi lowered its per-Token price, some subscribers still feel it is getting more expensive.

For model companies, Tokens are a cost that can be precisely measured; for users, what they really care about is how much work output a subscription can actually deliver.

When models begin to work on people's behalf, computing resources become one of the most important means of production in this era.

And China's large model startups have grown up precisely in an environment where such means of production are relatively scarce.

If we rewind to 2023, the large model startup waves in China and the United States emerged at almost the same time, but the starting points of these young companies were quite different.

In January 2023, Microsoft announced a new multi-billion-dollar investment in OpenAI, and their cooperation extended from funding to Azure cloud computing and supercomputing infrastructure.

In the same year, Anthropic established deep partnerships with Google and Amazon successively, obtaining massive funding and computing power support.

And tech giants such as Google and Meta already possessed enormous AI computing infrastructure themselves.

China's situation was much more complicated.

In October 2022, the United States began imposing export controls on advanced computing chips to China, and further tightened restrictions in 2023.

It was precisely during this period that the domestic large model startup wave began to rise. A group of companies such as Moonshot AI and Baichuan Intelligence entered the competition in frontier models, and existing teams such as Zhipu also accelerated their R&D pace.

Although these companies also had support from capital and large technology enterprises, and DeepSeek also possessed computing resources accumulated by High-Flyer over many years, Chinese model companies always faced more restrictions in their access to advanced computing power, costs, and supply certainty.

This is also why I have always felt that computing power is the wound of the original family for China's large model startups.

A company's origins often influence its development path for a long time.

For China's large model startups, constraints on computing resources have gradually permeated every link of technology R&D, product supply, and commercialization.

DeepSeek is a very representative example.

From the MoE architecture and MLA attention mechanism to a series of engineering innovations centered on training efficiency and inference costs, DeepSeek has consistently shown a high degree of emphasis on computational efficiency.

These technologies are also directions jointly explored by frontier models worldwide, but in an environment where advanced computing resources are relatively limited, how to make each computing card deliver greater value through architecture and engineering optimization clearly has more urgent practical significance.

This pursuit of efficiency has gradually become a distinct engineering culture among Chinese model companies, and has also helped them continuously approach the global technological frontier under limited computing budgets.

But the more difficult choice than the technology route occurs at the commercialization stage.

For an independent model company, limited computing power needs to simultaneously support R&D for the next-generation model and inference services for existing products.

Competition in frontier models is becoming increasingly fierce, and it is difficult to slow R&D investment. Once a new model succeeds, it will attract more users and further drive up inference demand.

For tech giants with large cloud businesses such as Alibaba and ByteDance, part of the pressure can be alleviated through existing infrastructure and diversified revenue.

But for independent model startups, every resource allocation may mean giving up another important opportunity.

What is crueler is that for these companies, technological success sometimes instead becomes an expensive burden.

If one day Kimi really builds the world's best Coding model and attracts millions of developers to use it every day, then how much computing resources would be needed to let these users run their Agents stably and continuously, and how enormous would the inference costs be?

This realistic pressure also helps us understand why open source has become such an important strategic choice for China's large model industry.

Once model weights are opened, cloud vendors, inference service providers, and enterprise users around the world can use their own infrastructure to deploy the models, allowing the model's influence to break through the limits of the original developer's computing scale while building a broader developer ecosystem.

Model companies are responsible for creating intelligence, while the entire ecosystem shares the cost of keeping intelligence running.

However, as more and more enterprises are able to deploy models and provide inference services themselves, how can the companies that initially bore enormous R&D costs obtain sufficient revenue from the technological value they created to sustain R&D for the next-generation models?

This may also be the most difficult part for China's independent model startups.

Over the past few years, they have gradually proven that even in an environment with limited resources, China can still build world-class large models.

But between training a world-class model and building a world-class model company capable of sustainable operation, there still lies enormous computing infrastructure and a business model that is still being explored.

Fortunately, over the past two years, progress in domestic computing power has been faster than many people imagined.

I myself used Huawei's Ascend 910 series chips a few years ago, and the experience at that time was really not pleasant.

Hardware performance was one aspect; what was more troublesome was that the entire software ecosystem was far from mature. From framework adaptation and operator support to various underlying tools, engineers needed to invest a great deal of effort in debugging.

Using domestic cards back then was like driving on a road that had not yet been repaired: after going a stretch, you had to stop and clear obstacles.

But the changes over the past few years have indeed been significant.

Huawei's Ascend chips have evolved from the early 910 series to the 950 series. According to officially published technical specifications, Ascend 950's FP8 computing power reaches 1 PFLOPS, while the 950DT to be launched in the second half of this year has 144GB of memory capacity and 4TB/s memory access bandwidth. Some hardware indicators are already comparable to Nvidia's H200 computing card.

Although actual training and inference efficiency still need to be tested under real workloads, such progress is already quite considerable.

At the same time, Cambricon continues to iterate its SiYuan series accelerator cards, and T-Head released the new-generation Zhenwu V900 in September this year, claiming that its computing capability reaches 3 times that of the previous generation M890. Domestic AI chips are gradually forming a more complete product system.

Huawei has also placed Ascend 960, 970, and even 980 on its product roadmap for the coming years, and has begun organizing large numbers of domestic AI chips into larger-scale computing systems through supernode interconnection.

Chip companies are busy increasing computing power supply, and model companies are also stepping up the construction of their own computing infrastructure.

In September this year, Bloomberg reported that DeepSeek planned to deploy at least 160,000 Huawei Ascend 950DT chips in a data center to be built in Inner Mongolia to support large-scale inference services.

Although this plan has not yet been fully publicly confirmed by both sides, and actual delivery also faces capacity constraints, its planned scale is already enough to reflect leading model companies' expectations for future computing demand.

Zhipu provides another very interesting case.

On September 17, Zhipu founder Tang Jie disclosed that they used an Infra Agent driven by GLM-5.3 to participate in building and optimizing the inference infrastructure for GLM-5.3-Flash.

According to officially published information, this system runs on a cluster composed of more than 100,000 domestic AI accelerator cards, and the Infra Agent raised end-to-end throughput to about 3 times the initial baseline in less than two weeks.

This case triggered a lot of discussion about RSI at the time, but I think it has another, more practical significance: Chinese large model companies have begun trying to use the intelligence they created to in turn improve the efficiency of computing infrastructure.

After all, when the supply of a computing card is limited, enabling it to handle more computing tasks through compilers, inference frameworks, and resource scheduling itself has very direct commercial value.

These advances are indeed encouraging, but I still believe that the computing power shortage in China's large model industry will be difficult to thoroughly solve for quite a long time to come.

Because what truly constrains computing power supply now involves the entire semiconductor industry chain.

Progress in chip design capability is certainly important, but transforming these designs into sufficiently large-scale, stable, and cost-controllable production capacity still requires joint progress in many links such as advanced processes, lithography equipment, high-bandwidth memory, and advanced packaging.

In September this year, Huawei publicly stated that domestic market demand for its AI computing equipment exceeded existing production capacity.

I had even previously heard a claim that Huawei's order schedule had extended to around 2030.

There is still a lack of publicly verifiable evidence for this specific timeline, but judging from the supply and demand situation that has been disclosed, the pressure on expanding domestic computing power production is indeed very great.

In my view, to truly alleviate this structural computing power shortage, China ultimately still needs breakthroughs in more foundational links such as advanced manufacturing equipment, processes, storage, and packaging, and this clearly requires more time than developing a new generation of chips.

Moreover, even if the supply capacity of domestic computing power continues to improve, the growth of model scale and the spread of Agent applications may still continuously create new computing demand.

This is like a race with no finish line.

Chip manufacturers are continuously expanding production capacity, model companies are stepping up the construction of clusters, and Infra teams are working hard to improve the utilization efficiency of each card. And every cost reduction may make more previously expensive AI applications feasible, thereby bringing a new round of demand growth.

We are creating computing power at an unprecedented speed, and we are also consuming computing power at an unprecedented speed.

For tech giants with massive capital reserves and mature cloud computing infrastructure, this may be a competition in which they can invest over the long term.

But for independent model companies that need to rely on continuous financing and limited commercial revenue to sustain R&D, every increase in demand may mean new infrastructure investment and greater financial pressure.

Over the past few years, Chinese model startups have already achieved remarkable technological progress under limited resource conditions, but to truly serve more users with these capabilities at a sustainable cost, a sufficiently powerful computing infrastructure is still needed.

At that time, Kimi was standing in the spotlight, and with its long-text capabilities it became one of the brightest stars in China's large model industry.

We went to discuss cooperation full of expectations, hoping to use its capabilities to build products for enterprise customers that had been difficult to achieve in the past, but in the end, because of insufficient computing power, we could only regretfully give up.

Looking back now, I have come to understand this matter somewhat more deeply.

In two years, China's large model industry has already traveled a quite difficult road.

Technological goals that were once out of reach are now being achieved one by one by a group of young Chinese researchers and engineers, and we finally have the opportunity to stand on the world's most cutting-edge stage and compete head-on with those tech giants that possess enormous resources.

I am very happy about these advances, and because of that I hope even more that they can truly be transformed into widely used products and services.

After all, when a group of people have gone through hardships and finally brought a model to world-class standards, yet still need to repeatedly weigh how to let more people use it, this is somewhat regrettable.

As for Kimi and this generation of Chinese large model startups, I hope that one day in the future they will finally be able to break free from this constraint, invest in R&D with greater ease, and welcome every customer willing to use their products with greater confidence.

I still remember those four words that Moonshot AI's co-founder told us at that dinner in 2024:

Not enough computing power.

Two years later today, Chinese large model companies have come a long way.

And I hope that the next time someone goes to them to discuss cooperation, they will finally no longer have to hear those four words.AI Venture Capital Daily Channel: Frontier Technology This content comes from the internet, original link. The views represent only the author and do not represent Huxiu's position. If there is a copyright issue, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change the world is on the Huxiu APP[ ]