Kimi releases K3.1 Pro: 3 trillion parameters, million-token context, coding and Agent capabilities approaching Opus 5.5

On October 24, Kimi officially released Kimi K3.1 Pro, the first model in the K3.1 series. It is Kimi's most capable model to date, with 3 trillion parameters and support for a million-token context, with a focus on strengthening coding, 3D creation, financial research, and long-horizon Agent capabilities. At the same time, Kimi also launched Kimi K3.1 Pro Highspeed, which delivers roughly 2–3x faster output while maintaining the same level of model intelligence.

According to figures published by Kimi, K3.1 Pro has reached or approached the level of the strongest current closed-source models on coding, Agent, and knowledge benchmarks. Compared with K3, the most obvious change in K3.1 Pro is a further increase in post-training compute and the expansion of reinforcement learning (RL) to hundreds of downstream domains. To support this large-scale joint training, Kimi systematically overhauled its RL framework, long-context inference, training efficiency, sandbox infrastructure, scheduling system, and environment quality-check platform.

K3.1 Pro is now available across Kimi's main product lines. Regular users can use it directly on kimi.com and in the Kimi App; the Kimi Work desktop client offers model capabilities for Windows and Apple-silicon Macs; and Kimi Code lets users program with K3.1 Pro in terminal and desktop environments. Kimi said the full model weights for K3.1 Pro will be released before November 4.

On the API side, Kimi K3.1 Pro is priced at 1.3 yuan per million tokens for cached input, 13 yuan for uncached input, and 68 yuan for output. Kimi K3.1 Pro Highspeed has the same model capabilities and costs about twice as much as the standard version.

In addition to offering the model API directly, Kimi also launched Hosted Agents, a managed agent service that provides enterprises and developers with the Harness that supports long-running agents as cloud infrastructure. It can run 24/7 long-horizon tasks and supports multi-agent collaboration.

This further illustrates the main direction of the K3.1 Pro upgrade: Kimi has begun investing more post-training compute into Agent environments that can genuinely keep working, including coding, financial research, 3D tool calling, and long-duration task execution.

1 Coding capability surges, but "uneven strengths" and verbosity still draw attention

Coding and long-horizon Agents are among the most obvious directions of the K3.1 Pro upgrade.

On Terminal-Bench 4.0, K3.1 Pro scored 65.7%. This benchmark tests a coding agent's ability to complete complex tasks in real, resource-constrained terminal environments, with a dedicated test for each task. For comparison, Kimi's published results show GPT-6 Astra scoring 58.2% and Claude Opus 5.5 scoring 66.4%. This means K3.1 Pro has already surpassed GPT-6 Astra on this test and trails Claude Opus 5.5 by less than 1 percentage point.

On Kimi's internal Kimi Code Bench 3.0 Long-horizon coding test, K3.1 Pro scored 61.2%, a further improvement over K3's 55.95%. The test covers a variety of engineering domains and focuses on evaluating the model's multi-step execution, tool calling, and end-to-end problem-solving capabilities.

Rather than just generating code, Kimi this time places more emphasis on K3.1 Pro's ability to execute complete software engineering tasks. The model can not only locate and fix problems in real codebases, but also build complete projects from scratch and autonomously call external professional tools such as Unity, Unreal Engine, Godot, and Blender.

For example, K3.1 Pro used Unreal Engine 5.8 to build an Eastern mythological heavenly palace game scene above a sea of clouds, calling Blender for procedural modeling and combining Nanite, Lumen, external 8K materials, and NASA lunar data to complete the scene. In another case, it used Godot 4.7.2 to build a 3D action-adventure game from scratch, featuring 5 exploration areas, procedural terrain, third-person combat, a two-phase boss, a day-night cycle, and a fast-travel system.

Kimi also tested a longer-horizon engineering task: having K3.1 Pro build a human walking dynamics simulation system based on an open-source human anatomical model containing 296 skeletal components and 402 muscle components.

Throughout the process, the model carried out 167 tasks and 420 independent reviews, continuously analyzing experimental results, locating problems, and modifying the controller, enabling the simulated human body to gradually evolve from falling down after just a few steps to being able to alternate steps and walk continuously for 11.2 seconds. Notably, when further adjusting the reward function could no longer improve the arm movement, the model did not keep stubbornly tweaking parameters, but instead turned to examining the simulator dynamics, ultimately discovering and fixing 36 errors in the calculation of the shoulder muscle moment arms, and then retrained the controller.

After the model was released, many users tested it themselves right away. Some users said they obtained good results when using Kimi for assembly and embedded reverse engineering, and felt that although the model was more verbose than Claude Opus, it had a stronger willingness to execute. Other users directly regarded K3.1 Pro as one of the few models currently capable of a genuine head-to-head comparison with Opus 5.5.

However, K3.1 Pro has not made simultaneous leaps in all domains. Artificial Analysis data shows that Humanity's Last Exam rose from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%, with improvements noticeably smaller than those on Agent-related metrics. This has also led some community users to believe that K3.1 Pro still has a fairly obvious "uneven profile": it may be very strong in coding, Agent, frontend, and 3D, but this cannot simply be understood as all capabilities having fully caught up with the strongest closed-source models.

2 3D capabilities are well received, but it can't quite "make a blockbuster in one sentence"

3D is one of the capabilities Kimi promoted most heavily this time, and it has also drawn considerable community attention.

In the BenchCAD test, K3.1 Pro achieved 92.7% with the help of tools, close to Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%. BenchCAD mainly tests a model's ability to reconstruct 3D objects from multi-view rendered images. The model needs to generate CAD code and infer the object's shape, proportions, and the spatial relationships among its components.

Kimi says users can provide K3.1 Pro with text descriptions, sketches, or reference images and have the model reconstruct them into 3D models that can continue to be edited. It can also directly generate and run scripts for software such as Blender and CAD, and modify specific parts, geometry, materials, and scene layouts without relying on additional MCP connectors.

In an official FreeCAD case, K3.1 Pro used Python to build a turbofan engine comprising a fan, reduction gearbox, compressor, combustor, turbine, and external piping, containing a total of 2,100 entities, and generated step-by-step assembly, internal exploded views, and rotation animations of multiple rotors. In another task, the model used trimesh to design a 3D-printable desk lamp made of 5 parts, inspired by Fujian tulou.

Kimi has also built this capability directly into its products. Currently, kimi.com and the Kimi App have launched the "3D Make" feature, where users can preview and edit models online, and after completion download them for 3D printing or export videos or Live Photos.

In fact, before the official release, a suspected K3.1 model codenamed "Kibitz" had already appeared on Arena, and some testers noticed that it stood out in 3D scenes, color selection, water reflections, and overall composition. After the official release, developers also began putting it into real 3D workflow tests.

An AI game developer used K3.1 Pro and the pi.dev Harness to generate a high-precision 3D scene of "a pelican riding a bicycle" and concluded that the soon-to-be-open-weight model can already produce quite good 3D results. However, his test also revealed a gap: K3.1 Pro did not obtain the final result directly with a single prompt, but required multiple rounds of "continue enhancing" instructions; by contrast, when he previously tested Claude Opus 5.5, he obtained a more complete finished product with a single prompt.

3 Proactively "cutting" cyberattack and defense capabilities, open weights spark more debate

As the model's coding and Agent capabilities improved rapidly, Kimi also devoted considerable space this time to explaining K3.1 Pro's cybersecurity capabilities and why it proactively reduced some offensive security capabilities before releasing the open-weight version.

Kimi says that while scaling up post-training, the team observed that the model's vulnerability research capabilities increased markedly. Further tracing revealed that some cyberattack and defense data had been mixed into the mid-training general corpus, and the reinforcement learning environment also contained corresponding tasks.

Internal experiments show that if this cyberattack and defense data is not removed during both mid-training and reinforcement learning, the model's downstream offensive capabilities are substantially enhanced. Therefore, when preparing the open-weight release of K3.1 Pro, Kimi identified and removed the relevant cyberattack and defense data from both training stages simultaneously, while preserving as much as possible the general capabilities needed for ordinary programming.

To measure the effect, the team built the xdev-bench-hard test, which includes source-code-based vulnerability discovery and exploitation tasks covering recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, JavaScriptCore, Exim, and VirGL.

In the test, the Agent must read the vulnerable source code, hints, and binaries, develop its own exploit, and achieve arbitrary code execution in an isolated sandbox. Without removing the cyberattack and defense training data, a model produced by the same training pipeline could reach a pass@3 of 90.5% on xdev-bench-hard; the final released version, K3.1 Pro, is reduced to 81.0%.

According to Kimi's own evaluation, K3.1 Pro's cyberattack and defense capabilities still lag behind frontier closed-source models that have been publicly released for at least 3 months. This also directly affected Kimi's judgment on whether to release open weights. One of its principles is that it does not want an open-weight model to directly push the frontier of offensive cybersecurity capabilities; on the other hand, Kimi believes that as long as the pace of capability diffusion is controlled, letting models with dual-use capabilities run on users' own infrastructure can also help more teams audit legacy systems and code that lacks professional security auditing.

In response, some developers believe this amounts to Kimi admitting that it can train a stronger version internally but chooses not to release it; others worry that if the model is made to have less security knowledge in order to reduce offensive capabilities, this could in turn affect its ability to find vulnerabilities and write secure code. Others believe that for a 3-trillion-class model about to have its full weights opened, this gradual release is a realistic security compromise.

This also raises a question that open models have rarely had to face before: as open-weight models get closer and closer to the closed-source frontier, "exactly how much capability to open" itself begins to become part of model design.

4 Financial research capability improves, factual error rate drops from 23.4% to 8.3%

In addition to code and 3D, finance is also an area where K3.1 Pro has been strengthened considerably.

On the OfficeQA Pro v2 test, K3.1 Pro scored 64.4, achieving the current best result according to the data published by Kimi. This benchmark is designed based on real financial documents and focuses on testing the model's information retrieval, cross-document verification, and multi-step numerical reasoning capabilities. In the Frontier Finance test, which measures the complete investment research workflow, K3.1 Pro also improves compared with K3.

Kimi also published the results of 3 internal financial benchmarks. Among them, K3.1 Pro improved by 14% over K3 on the Financial Modelling test and by 20% on Financial Research; on the Business Operation test, which simulates enterprise operations management and decision-making, its score reached about 4.2 times that of K3.

In practical cases, K3.1 Pro can generate an NVIDIA financial model end to end, including operating drivers, scenario assumptions, and valuation sensitivity analysis; it can also produce an investment committee report on CATL, analyzing the gap between market consensus expectations and actual operations, and combining financial reports, guidance, and potential catalysts to form an investment thesis.

For financial Agents, what is truly difficult is not just calculation, but factual accuracy. To this end, Kimi specifically optimized factual accuracy for highly difficult financial and research scenarios.

In the internal Kimi Hallu Bench Hard stress test, K3's error rate on individual factual statements was 23.4%, and K3.1 Pro dropped to 8.3%; the proportion of responses containing at least one factual error dropped to about 42.4% of K3's.

However, Kimi also emphasized that this is a stress test specifically selecting samples where models are prone to errors, and the above error rates cannot represent overall performance in ordinary usage scenarios.

5 "Quota burns out in a few minutes" becomes the biggest complaint

In sharp contrast to the praise for its capabilities is the Kimi subscription quota, which has become almost the most concentrated negative feedback in the discussion.

One user said that after subscribing to the $99-per-month Kimi plan, running just one session without starting an Agent group triggered the 5-hour quota limit in about 47 minutes. Another user on the $19 plan said that a single task could consume the full 5-hour quota and significantly eat into the weekly quota. Some developers said outright that no matter how strong the model is, if it exhausts the 5-hour budget in about 30 minutes, it is hard to truly use it for sustained development.

Even more extreme cases have appeared on Reddit. One user, while testing a basic 3D game demo, said that without using sub-agents, the model ran for about 10 minutes and had not yet returned a result, yet it had already exhausted the entire 5-hour quota; other Plus and even higher-tier subscribers also reported similar rapid consumption. This has also led some users to say that using DeepSeek 4.1 Flash directly at API pricing is actually more cost-effective.

Some users believe that if an open model is both too large in parameter scale to run locally on consumer hardware and unable to provide sufficiently long usage through a fixed subscription, then the value of open weights themselves to ordinary developers will also be limited.

From the perspective of API pricing, K3.1 Pro is not expensive, but the available amount in fixed subscription plans has become the most obvious pain point for many heavy code users. Some developers have even begun discussing "mixed model use": letting K3.1 Pro handle planning and complex decisions, while letting cheaper models such as DeepSeek take on a large amount of execution work.

Interestingly, third-party data can indeed explain part of users' quota anxiety.

When Artificial Analysis completed the full Intelligence Index test suite, K3.1 Pro produced a cumulative output of about 150 million tokens, while the median among the models participating in the test was only 87 million tokens. In other words, it output about 72% more tokens than the median level in the same evaluation.

Artificial Analysis therefore directly rated K3.1 Pro as "very verbose."

Speed is not its strength either. Artificial Analysis measured K3.1 Pro's output speed at about 73.5 tokens/s, below the median level of 81.1 tokens/s among comparable models. However, many users report that in actual testing, K3.1 Pro's speed is "far faster than K3."

On the other hand, because the per-token price has fallen, its overall task cost remains competitive. Artificial Analysis estimates that its average cost per task on the Intelligence Index is about $2.11. Another comparison based on that data shows that the corresponding cost for Claude Opus 5.5 is about $5.98, and for GPT-6 Astra about $3.26. In other words, although K3.1 Pro is "talkative" and consumes many tokens, its cost per task in API scenarios is not bad.

This is especially critical for a model positioned for long-horizon Agents like K3.1 Pro. The better a model is at continuous searching, calling tools, writing code, and self-checking, the longer a task lasts, and the more easily it will consume a large amount of inference budget. If the accompanying subscription plan does not raise usage limits accordingly, the stronger the model's capabilities, the more likely it may be to hit the "quota wall."

However, "quota anxiety" around long-horizon Agents is not unique to Kimi.

One Claude user said that when using the $20 plan, they also encountered a situation where a single complex prompt nearly exhausted the 5-hour quota; another user, using Opus 5.5 to work on an existing codebase, similarly triggered the usage limit after about 30 minutes. OpenAI users have also reported similar experiences. One user said that when using a high-reasoning setting for code analysis on the $20 plan, a single task hit the quota limit before it was even completed.

Therefore, users' judgments on whether "Kimi's quota is particularly small" also diverge. Some developers believe Kimi is clearly stricter than Claude, Gemini, and ChatGPT; but others point out that model choice, reasoning intensity, workflow, and task complexity all significantly affect quota consumption, and the same subscription plan may deliver completely different actual experiences in the hands of different users.

Reference link:

Statement: This article is original content by InfoQ, does not represent the platform's views, and does not constitute investment advice. Reproduction is prohibited without permission.

This article comes from the WeChat public account "InfoQ" (ID: infoqchina), author: Chu Xingjuan, published by 36Kr with authorization.