Kimi K3.1 Pro Released: Capabilities Improved, but Price Is 16 Times That of DeepSeek and MiniMax Flash

[[**Kimi releases K3.1 Pro, boosted to evaluation leadership through post-training, but priced far above DeepSeek and MiniMax Flash; under the dual goals of commercialization and SOTA, balancing resources, cost, and experience comes under pressure.**

**Key points**

**1. Capabilities improved without scaling parameters**

Maintains 2.8 trillion parameters, no pre-training redo, increased post-training compute and joint reinforcement learning across hundreds of domains. AA composite index about 55 points, up 11 from about 44 for K3, leading the two Flash models.

**2. High capability comes with a high price**

Charges $2/$10 per million input/output tokens; at 1 million each that is $12, 16 times the $0.75 off-peak rate of MiniMax and DeepSeek and 8 times DeepSeek's peak rate, far above Flash.

**3. The three companies' commercialization focuses diverge**

Kimi strengthens 3D modeling and financial research, DeepSeek emphasizes full project development, and MiniMax focuses on computer operation and cross-application workflows.

**4. Under dual goals, experience and renewals come under pressure**

ARR is reportedly set to exceed $1 billion by August 2026, with a path to $2 billion by year-end, and overseas revenue higher than domestic. Kimi plans to continue scaling parameters in its next-generation flagship while also developing Flash; however, the high token consumption of 3D tasks is already affecting user experience and renewals. Original 2026-10-24 20:41 Kimi's biggest paradox is that it wants everything, but compute is finite. Song Sihang Quick Overview

In the early hours of October 24, Kimi K3.1 Pro was released. Moonshot AI's official WeChat account wrote, "We are officially launching Kimi K3.1 Pro, the first model in the K3.1 series and our most capable 3-trillion-level parameter model to date." It has been 99 days since Kimi last released the K3 flagship model.

It is worth noting that the "3-trillion-level parameters" here does not mean the model has expanded its parameters again. K3.1 Pro still maintains K3's 2.8T, that is, a total parameter scale of 2.8 trillion. This upgrade did not involve retraining from scratch; the focus was on further strengthening post-training on top of the existing base. The official explanation also made it clear that K3.1 Pro is a post-trained version of K3, with a substantial increase in post-training compute investment, especially joint reinforcement learning across hundreds of downstream domains.

And in the past two weeks, domestic model vendors have also successively released their latest models: first, on October 17, DeepSeek released V4.1 Flash-1017, and then, on October 19, MiniMax released M3.1 Flash. Both emphasized higher inference efficiency and lower task costs while improving capabilities.

By contrast, Kimi's focus this time is somewhat different. Although the API price of K3.1 Pro has also dropped compared with K3, it still sits in a clearly higher price range than the Flash models DeepSeek and MiniMax launched this time. What Kimi launched first this time is the Pro version, which continues to raise the capability ceiling.

But at the same time, Huxiu also learned that Kimi is developing a Flash version and plans to continue expanding the parameter scale in its next-generation flagship model in pursuit of SOTA.

This means that at a key juncture before its IPO, Moonshot AI has both maintained its previous approach of continuing to pursue SOTA and also sought to improve its commercialization results through Flash versions, Pro versions, and so on. But the question is: within a limited pool of resources, how can it balance the two?

According to Bloomberg, as of August 2026, the ARR Moonshot AI disclosed to investors had surpassed $1 billion, up from $300 million in June, and by the end of this year its ARR is expected to surpass $2 billion. Notably, when Kimi revealed its commercialization progress earlier this year, it emphasized that its overseas revenue was far higher than its domestic revenue.

The release of the K3.1 Pro version this time reflects this: while domestic model vendors are all pursuing cost-effectiveness, Kimi has continued to choose to push the upper limit of capability, which is of course a good thing. However, based on feedback from some developers so far, K3.1 Pro is not perfect either. Some users report that when using the 3D-Maker feature, a single task consumes too many tokens and causes them to hit the quota limit, affecting the user experience.

After comparing with DeepSeek V4.1-Flash and MiniMax M3.1-Flash, it was found that K3.1-Pro is 16 times that of the former two.

Looking back at the models Kimi has released in the past, although there have been versions that emphasized speed and specific tasks, Flash had not previously become a major product naming convention, and this time is no different. For a model company that is pursuing commercialization results, products that balance speed, cost, and capability are becoming increasingly important. In this round of competition over Flash products, Kimi has yet to make its move.

But Kimi's model approach this time is somewhat different. Flash usually emphasizes a balance of speed, cost, and capability, while Pro leans more toward complex tasks and the upper limit of capability.

In this regard, Huxiu compared the three recently released models. Although Kimi K3.1 Pro, DeepSeek V4.1 Flash-1017, and MiniMax M3.1 Flash do not belong to the same tier, they all target programming, tool calling, and long-workflow tasks, and in these shared scenarios they are still comparable.

First, let's look at capability. In the comprehensive intelligence index of third-party evaluator Artificial Analysis, Kimi K3.1 Pro scores about 55, DeepSeek V4.1 Flash-1017 scores about 45 under the max reasoning setting, and MiniMax M3.1 Flash scores about 43. At least within this evaluation system, Kimi leads in overall capability, while the scores of the other two Flash models are fairly close.

Source: Artificial Analysis

More noteworthy is that, compared with K3's score of about 44, K3.1 Pro improved by about 11 points. This means that even without further expanding the parameter scale, Kimi still achieved a clear capability gain through this round of updates. According to AA's evaluation, this progress is more prominent on agent tasks, while the improvements on some knowledge and scientific reasoning tests are relatively modest.

However, higher capability also corresponds to a higher calling price.

Under the three official USD API prices, Kimi K3.1 Pro charges $2 per million uncached input tokens and $10 for output; MiniMax M3.1 Flash charges $0.15 and $0.60 respectively, and DeepSeek V4.1 Flash-1017 matches MiniMax during off-peak hours and doubles during peak hours.

Calculating on the basis of 1 million cumulative input tokens and 1 million output tokens, without using caching and excluding other costs such as tool calls, Kimi's bill is $12, while MiniMax and DeepSeek during off-peak hours are both $0.75, making Kimi's cost 16 times that of the latter two; relative to DeepSeek's $1.50 during peak hours, it is 8 times. Compared with K3, the input and output unit prices of K3.1 Pro have already dropped by about one-third, but even after the price cut it remains significantly higher than the other two Flash models.

Of course, this gap does not necessarily mean that Kimi's cost-effectiveness is lower. Equal token usage does not equal equal work output. For work that all three models can complete, a lower price is more attractive; but if Kimi can solve problems that the other two models find difficult to complete, or significantly reduce manual rework, a higher price may be justified.

Therefore, what is more worth further observation than the total score is what work capabilities each of the three is strengthening, and what kind of paid demand these capabilities can attract. At the same time, the three companies' key directions in RL also reflect their different directions of commercial exploration.

Judging from the focus of this release, Kimi places more emphasis on 3D modeling and financial research, DeepSeek emphasizes complete project development, and MiniMax puts computer operation and cross-application workflows in a prominent position.

The 3D and financial scenarios chosen by Kimi both require the model to be combined with professional tools and expertise. For example, 3D modeling requires converting the shape and spatial relationships of objects into usable model assets; financial research not only requires organizing information, but also involves cross-document verification, numerical reasoning, and financial modeling. Compared with simply generating a passage of text, this places more specific requirements on the final result and is also more likely to expose shortcomings in the model's capabilities.

This emphasis is already reflected in the product. At this release, Kimi simultaneously launched the 3D Make feature, which supports online model preview, editing, and export for 3D printing. In the financial direction, the internal tests it disclosed separately assess financial modeling, research, and simulated business operations capabilities. Although the internal results still need more external validation, it is clear that Kimi is trying to translate improvements in general capabilities into usability in professional scenarios.

By contrast, what DeepSeek and MiniMax emphasized in this round is enabling execution capabilities in development and office work to be used at lower cost. The former demonstrated complete projects such as developing a game from scratch and building 3D scenes, while the latter emphasized software development, computer operation, and cross-application work in sales, operations, finance, and other areas.

DeepSeek's 1017 version uses self-feedback reinforcement learning to improve project development capabilities while also shortening task duration and output length; its V4.1 Flash base had previously already reduced the burden of long-context processing through compression, KV cache reuse, and other methods. MiniMax, meanwhile, adopts more aggressive sparsification and claims that M3.1 Flash's online inference speed is nearly three times that of the previous generation, with the overall cost of Agent tasks dropping by 70%.

Kimi, by contrast, explicitly increased its post-training compute investment, conducting joint reinforcement learning across hundreds of downstream domains, along with supporting infrastructure such as optimized training frameworks, sandboxes, and task scheduling. Its more notable choice this round was to continue tapping the potential of its large base model on high-difficulty tasks, and then use real-world results to prove the value of Pro.

Kimi wants both commercialization and SOTA

Judging from the publicly disclosed parameter scale, Kimi K3 is already among the largest domestic models. But for its next-generation flagship model, Kimi is still aiming for a larger-parameter SOTA model. All else being roughly equal, this means higher training compute requirements.

Unlike before, Kimi is also accelerating commercialization this year, and overseas markets have already become an important source of revenue. According to a report by The Paper in February this year, Kimi's cumulative revenue over the preceding nearly 20 days had already exceeded that of the whole of 2025, and overseas revenue had also surpassed domestic revenue. By August, according to Bloomberg, the ARR disclosed by Moonshot AI to investors had exceeded $1 billion, up from $300 million in June, and by the end of this year its ARR is expected to exceed $2 billion.

This also gives Kimi's choice to pursue SOTA a more direct commercial explanation. For overseas markets, what it needs to win over are users who are already using overseas flagship models. In the official evaluation of K3.1 Pro, Claude Opus 5.5 and GPT-6 Astra are the main comparison targets. Strengthening professional capabilities such as coding, 3D modeling, and financial research can also be understood as a way to persuade this group of users to pay.

However, while Kimi keeps pushing the upper limit of model capabilities, user experience is equally important. After K3.1 Pro launched, a user on Reddit reported that they asked the model to create a basic 3D game demo without enabling sub-agents, and it exhausted the usage quota for the five-hour window in about 10 minutes, still without getting a result. This means that 3D tasks may consume more tokens, thereby reaching the quota limit and ultimately affecting user experience.

If complex tasks frequently hit the quota limit, then the capabilities the model demonstrates will be hard to fully translate into value in users' everyday use. For Kimi, which hopes to retain overseas paying users, task completion quality, quota consumption, and service continuity will all affect the next renewal.

Therefore, what Moonshot AI needs to balance is not only the training investment for the next-generation flagship, but also the inference costs and user experience of existing products. This is also the meaning of the Flash version: use lower-cost models to handle everyday tasks, and let the flagship serve more complex needs. However, the premise of lowering costs is still being able to complete tasks; otherwise, the saved invocation fees may be offset by retries and manual rework.

Commercialization and SOTA are not inherently in conflict. But for Kimi, it still needs to find a balance and make some trade-offs.

Article title: Kimi's biggest myth is that it wants everything, but compute is limited

Has published 95 articles on Huxiu. This content may not be reproduced without permission. For licensing matters, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]