Kimi K3.1 Pro apologizes for quota consumption, contrasting with DeepSeek's token-saving approach

Less than 48 hours after the release of K3.1 Pro, Moonshot AI issued an announcement saying, "We sincerely apologize for the rate limiting and the unexplained depletion of quotas in the model. These problems are entirely our responsibility, and we should have pushed to resolve them sooner. Starting October 26, all subscribers across every tier will receive a refund of 25% of their monthly total. We have also adjusted the product's operating parameters, cutting unnecessary sub-Agent calls and compressing redundant output length to make quotas last longer."

After this change, the most noticeable difference in experience is that K3.1 Pro no longer "burns through quotas" the way it did right after launch.

Going back to the day the model was released, the entire AI community gasped: its performance surpassed all OpenAI models, trailing Opus 5.5 by only half a foot. The third-party evaluation firm ModelCompare wrote that the AI gap between China and the US had shrunk from 3 to 4 months to just 31 days.

Yet this celebration did not even last two days before K3.1 Pro's rating went from "Overwhelmingly Positive" to "Mostly Positive." Yang Zhilin was also given the nickname "dual nature of god and ghost"—god because the model's performance broke through the sky, ghost because using K3.1 Pro was somewhat "painful to the wallet."

The most frequent feedback was about quotas. Multiple users in the community said they had barely used it before exhausting K3.1 Pro's 5-hour quota. Take myself as an example: on the 199-yuan tier, in two days I generated fewer than 10 3D models and a few demo animations no longer than 30 seconds, and half of my monthly quota was already gone.

Even for some simple tasks, K3.1 Pro calls a large number of sub-Agents to think and reason. Its output results are not much improved, yet because of these sub-Agents, users waste a lot of quota.

Yang Zhilin is rather helpless about this too: raising a model's intelligence level inevitably requires more tokens.

This also means that although K3.1 Pro is 30% cheaper than K3, because the quotas given are too small, in the end not many people truly felt the "cheapness."

Just one week before K3.1 Pro's release, DeepSeek released V4.1 Flash-1017, whose approach was completely the opposite: rather than competing on the upper limit, it optimized performance, speed, and output length together while keeping the price unchanged.

V4.1 Flash-1017 perfectly embodies Liang Wenfeng's product philosophy, and since its release it has remained "flooded with praise."

Perhaps Yang Zhilin should learn a thing or two from Liang Wenfeng when it comes to products.

Quotas are Moonshot AI's biggest wound

After K3.1 Pro's release, many people in the community discussed the quota issue. For example, there was a highly upvoted comment in which the poster said, "It took it only about 10 minutes, and the 5-hour quota window was already full."

On Linux.do, a user on the 199 plan said that he only sent one instruction asking the model to review the project architecture, and Kimi spun up 5 sub-Agents; more than ten minutes later, the 5-hour quota was full.

English text is being prepared.

English text is being prepared.

Artificial Analysis believes that the key reason K3.1 Pro consumes quotas so heavily lies in token efficiency.

In testing, it found that running the entire Intelligence Index test suite on K3.1 Pro requires outputting about 150 million tokens, while the median for models at the same price point is 87 million.

On the same test suite, GPT-6.1 Sol used only about 67 million, and GPT-6 Astra about 60 million. Converted according to OrcaRouter, K3.1 Pro's output for the same task is about 2.5 times that of Astra.

Artificial Analysis also said that K3.1 Pro's speed is below the median for its peers, at roughly 70 to 80 tokens per second. In addition, K3.1 Pro only supports thinking mode; requests with thinking turned off are directly rejected, so every call pays for reasoning.

Although K3.1 Pro has launched a Highspeed version, Moonshot AI officially stated that the Highspeed version only speeds up model output. The time consumed by tool calls in a task (reading and writing files, executing commands, etc.) and script execution is unaffected; if these account for a large proportion, the overall speedup will not be obvious.

Although K3.1 Pro is 30% cheaper than K3, Artificial Analysis found in testing that K3.1 Pro's per-task cost is not much cheaper.

K3.1 Pro's average cost per task is $2.11, while GPT-6 Astra's is $3.26. K3.1 Pro's token price is only one-fifth of Astra's, yet in terms of actual spending, it saves only about 35%.

All of this may be related to K3.1 Pro's training.

Yang Zhilin believes that to enable a model to produce longer outputs, it must remain rigorous throughout the entire process, repeatedly checking, revising, and self-correcting. This has been the case since K3; the official technical documentation at the time stated that K3 would exhibit excessive caution and repeated self-scrutiny in its outputs.

The official Fireworks AI research blog "Introducing Ember-1" mentioned that even removing K3's repeated checking process had almost no impact on model performance. In other words, K3 is a bit too rigorous.

By K3.1 Pro, the model had to be rigorous not only about the task but about the entire environment.

K3.1 Pro's official technical report mentioned a case in which the model simulated a person learning to walk. K3.1 Pro found that the arm movements were consistently unsatisfactory and that adjusting the reward did not solve the problem, so it went back to check the simulator, fixed 36 calculation errors in the shoulder muscle moment arms, and then retrained, ultimately achieving better results.

This is also why K3.1 Pro, despite its high performance, produces longer outputs and more tokens: it has to remain rigorous about almost everything, even the environment.

Yang Zhilin burns tokens, while Liang Wenfeng saves tokens.

DeepSeek called the V4.1 Flash 1017 iteration a "Pareto improvement." This is an economics concept referring to a change after which at least one aspect improves while no aspect worsens.

In the context of a model, this means higher scores across benchmarks, faster task completion, shorter outputs, and no change in price.

Someone in the DEV community ran tests and said that compared with 4.1 Flash, DeepSeek V4.1 Flash 1017 produces shorter outputs, completes tasks faster, and actually lowers the cost per completed task.

In fact, V4.1 Flash itself is an extremely "chatty" model. A real-world test on OrcaRouter showed that on the same Artificial Analysis Intelligence Index, V4.1 Flash has to output 250 million tokens, 56% more than K3's 160 million.

However, thanks to V4.1 Flash's Causal Encoder-Decoder architecture, with 552B total parameters but only 8B activated for prefill and 16B for decode, the KV cache is compressed to only 890 bytes per token, VRAM usage drops to a quarter of the previous generation, and SSD usage drops to an eighth.

So DeepSeek scores higher, completes tasks faster, and even shortens outputs; Liang Wenfeng is unwilling to sacrifice one attribute for another.

Agents are one link in the service of AGI

Liang Wenfeng once said, "What investors see most now is Agents, but for those of us doing research, what we see more now is learning, and how to solve the problem of learning."

The implication is that Liang Wenfeng believes Agents are only one link in AGI, not the whole of it.

In 2025, Yang Zhilin appeared on Zhang Xiaojun's "Business Interview." Regarding pure reasoning models and Agents, Yang Zhilin proposed a brain-in-a-vat theory: "Imagine a fish tank, and you put a brain inside it, with no connection to the outside world. It just thinks inside its own brain, keeps thinking, and can solve a problem without needing any interaction with the outside world."

Regarding Agents, Yang Zhilin said: "It is no longer a 'brain in a vat'; it interacts with the outside world. My next action depends on the feedback obtained through interaction and on updates to the new state the outside world gives me."

The brain in a vat represents strong reasoning models like o1, but what Yang Zhilin wants is for models to be able to use more tools and engage in more rounds of interaction.

However, no matter what Liang Wenfeng and Yang Zhilin say, that is their definition of Agents in terms of product philosophy. To really see how a model's Agent capabilities are, you still have to put it into real tasks.

Besides Artificial Analysis, the community also often uses the "ZhiZhi | Long-term Tracking Leaderboard of Large Model Intelligence" (Cat Leaderboard) to test models' levels.

It breaks model capabilities into three sections: reasoning, Agentic, and vision. On the Agentic page, it uses a set of tasks requiring long workflows, multiple constraints, and tool collaboration. The numbers and letters on the left indicate penalty points and tiers; the page says "1/A represents penalty points/tier," and the smaller the number, the better. The ￥ amount on the right appears to be the cost for the model to complete that problem.

Compared with the previous generation K3, K3.1 Pro's progress is very obvious. On Rust App, K3 was "Failed," meaning it could not implement it, while K3.1 Pro scored 23 points, tier C; on Science, K3 scored 26 points, C+, and K3.1 Pro improved to 16 points, tier B; Animation also improved from 22 points to 17 points.

This result is consistent with what Moonshot AI officially emphasized: "Agent execution capability is greatly improved."

However, on this leaderboard, K3.1 Pro lost to DeepSeek V4.1 Flash 1017.

On Rust App, DeepSeek V4.1 scored 12 points, tier B; K3.1 Pro scored 23 points, tier C; on the Simple Model item, K3.1 Pro scored 5 points and DeepSeek scored 10, both in tier B+, with Kimi slightly better; on the Animation item, both scored 17 points, tier B, a tie; on the Science item, DeepSeek scored 13 points and Kimi scored 16, both in tier B, with DeepSeek slightly better.

In addition, K3.1 Pro is still "Pending" on the Data Process and Metal items, meaning a complete conclusion cannot yet be drawn.

K3.1 Pro, which clearly rivals Opus 5.5 on Artificial Analysis, scores lower than DeepSeek here?

This is because the Mao leaderboard has a rule: "The model is not limited on Max Token, but ultra-long thinking is penalized with point deductions based on question difficulty." This means the more tokens you spend, the more points are deducted, and K3.1 Pro suffers a big disadvantage.

On Rust App, DeepSeek spent 12.32 yuan and K3.1 Pro spent 221.21 yuan. On Simple Model, Kimi spent 12.60 yuan and DeepSeek spent 5.88 yuan. On Animation, Kimi spent 38.77 yuan and DeepSeek spent 10.77 yuan. On Science, Kimi spent 25.80 yuan and DeepSeek spent 6.21 yuan.

There is another point: K3.1 Pro's technical blog states that the model will exhibit "excessive proactiveness," making decisions for you when it encounters ambiguous intent and being overly improvisational. Especially for Agents, K3.1 Pro is more willing to freely improvise based on the prompt.

But in the Mao leaderboard, tasks like Rust App have extremely long environment feedback chains and completely fixed requirement boundaries, which strictly limit the model's room to perform.

What Liang Wenfeng is doing is making every Agent call stable and cheap; what Yang Zhilin is doing is making the hardest few scenarios look impressive enough.

Moonshot AI needs "restraint"

K3.1 Pro's 3D capability is one of the highlights of the release. You can give the model a few images without explaining anything, and it can convert them into 3D models.

The day after the release, a highly upvoted post appeared in the community. The poster said he discovered that Kimi has a built-in skill called model3d_generation, which runs a Python script to generate 3D models through an external API service.

The poster deliberately disabled this skill and had K3.1 Pro generate only with its own code, and the quality was noticeably worse. More importantly, the script supports three model versions: v3.1, v3.0, and v2.5, which exactly match the model versions of the 3D generation company Tripo AI.

In fact, even before K3.1 Pro was released, the Kimi Work desktop client had already included Tripo 3D in its plugin ecosystem.

An article titled "How to Create 3D Image Designs with Realistic Effects," updated on October 17 in the Moonshot AI Resource Center, explicitly states that Kimi Work "supports Blender, Tripo 3D, and other plugins."

The official blog post from October 14, "10 Free and Paid AI 3D Image Generators Worth Trying," also mentions that Kimi Work has integrated Tripo 3D.

In fact, Moonshot AI's own technical blog explicitly states that when K3.1 Pro works in a game engine, it will "combine web search and 3D generation" to create a complete experience. The 92.7% on BenchCAD is also labeled "with tools."

This means the official side did not say these results were achieved by the model "bare-handed."

There is nothing wrong with asking for "outside help." Calling tools is originally part of Agent capability; a model that can find suitable tools and integrate them into its workflow is more useful than one that brute-forces everything itself.

But the problem is that Yang Zhilin wants too much. In addition to 3D modeling and game engines, K3.1 Pro also covers code, long-horizon Agents, financial research, and so on.

Each direction, viewed on its own, makes sense and has real progress.

Artificial Analysis data shows that K3.1 Pro scores 34.3% on scientific reasoning, while Opus 5.5 is at 59.0% and GPT-6 Astra at 63.3% over the same period. On the HLE test without tools, K3.1 Pro is at 51.1%, lower than Opus 5.5's 59.8%; long-context reasoning also slipped slightly from K3's 88.7% to 86.3%.

The front is stretched too long, so the resources each line can get become fewer.

By contrast, Liang Wenfeng is very restrained. He once said: "This is also part of our restraint. I hope these business opportunities can be left to others. We hope these business opportunities and how to use this AI can be done by the whole society and all our partners together, and everyone can share the benefits together, rather than me wanting to monopolize them. That is impossible."

DeepSeek is not doing nothing; it also has search and scraping APIs and various productization moves. The difference is that everything is placed under the premise of "serving the AGI main line." Things that do not serve the main line, even if they are good business, are preferably not done.

Restraint is not indifference; it is essentially a method of resource allocation. First decide what not to do, then decide how to do it well.