International access: https://4sapi.com
Domestic access: https://4sapi.org
Abstract
On October 24, Moonshot AI officially released Kimi K3.1 Pro, the first model of the K3.1 series and the strongest model Kimi has produced to date. The model belongs to the 3-trillion-parameter class, supports a one-million-token context window, and strengthens code, 3D creation, financial research and long-horizon agent capability. A Highspeed variant was released simultaneously. Community reaction is sharply divided: strong praise for code, 3D and agent performance collides with concentrated complaints about rapid subscription quota consumption. Third-party data from Artificial Analysis shows that K3.1 Pro emits roughly 72% more tokens than the median model in identical evaluations, which directly explains much of the quota anxiety. This article organizes the benchmark figures, real-world engineering cases, the deliberate reduction of cyber-offensive capability before weight release, financial research gains, and the token-consumption debate.
1. Release Overview and Core Benchmark Results
K3.1 Pro is positioned as a full-stack engineering model rather than a chat assistant. Moonshot AI explicitly strengthened the model's ability to execute complete software engineering tasks, not just generate code snippets. The model can locate and fix issues in real codebases, build complete projects from scratch, and autonomously invoke external professional tools including Unity, Unreal Engine, Godot and Blender.
On Terminal-Bench 4.0, K3.1 Pro scored 65.7%. This benchmark tests coding agents inside real, resource-constrained terminal environments, with dedicated tests for every task. Moonshot published comparison figures showing GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%. K3.1 Pro therefore exceeds GPT-6 Astra on this test and sits within one percentage point of Claude Opus 5.5.
On the internal Kimi Code Bench 3.0 Long-horizon test, K3.1 Pro reached 61.2%, up from K3's 55.95%. This suite spans multiple engineering domains and focuses on multi-step execution, tool invocation and end-to-end problem solving. Both results indicate that the model's strongest improvements concentrate in agentic and engineering workflows rather than general conversation.
2. Real-World Engineering Demonstration Cases
Moonshot presented two notable 3D development cases to demonstrate end-to-end capability. In the first case, K3.1 Pro used Unreal Engine 5.8 to construct a Chinese-mythology heavenly-palace game scene above a sea of clouds. It invoked Blender for procedural modeling and combined Nanite, Lumen, external 8K textures and NASA lunar data to complete the scene.
In the second case, the model built a full 3D action-adventure game from scratch using Godot 4.7.2. The output contains five explorable regions, procedural terrain, third-person combat, a two-phase boss battle, a day-night cycle and a fast-travel system. These examples demonstrate coordinated tool orchestration rather than isolated code generation.
The most demanding long-horizon engineering test involved building a human-walking dynamics simulation system. The model worked from an open-source human anatomy model containing 296 skeletal components and 402 muscle components. Across the entire process, it executed 167 tasks and 420 independent reviews, continuously analyzing experimental results, locating problems and adjusting controllers. The simulated human progressed from falling within the first few steps to alternating strides and continuous walking for 11.2 seconds. When tuning the reward function could no longer improve arm movement, the model stopped iterating on parameters and redirected attention to the simulator dynamics. It discovered and fixed 36 errors in the shoulder muscle moment-arm calculation, then retrained the controller.
3. User Feedback and the "Lopsided" Capability Profile
After release, many users tested the model independently. One user reported strong results in assembly and embedded reverse engineering, noting that the model is more verbose than Claude Opus but shows a stronger tendency to execute. Other users framed K3.1 Pro as one of the few current models that can be compared head-to-head with Opus 5.5.
Not every domain improves at the same rate. Artificial Analysis data shows Humanity's Last Exam rising from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. These gains are notably smaller than the agent-related metric improvements. This has led part of the community to conclude that K3.1 Pro remains visibly lopsided: code, agent, frontend and 3D may be very strong, but it cannot be assumed that every capability has caught up with the strongest closed-source models.
4. 3D Capability: Widely Praised but Not "One-Prompt Cinema"
3D is the capability Moonshot promoted most aggressively, and it drew heavy community attention. On the BenchCAD test, K3.1 Pro scored 92.7% with tool assistance, close to Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%. BenchCAD tests a model's ability to reconstruct 3D objects from multiple rendered views. The model must generate CAD code and infer object shape, proportions and spatial relationships between components.
Users can provide text descriptions, sketches or reference images, and the model rebuilds them into editable 3D models. It can also directly generate and run Blender and CAD software scripts, modifying specific parts, geometry, materials and scene layout without relying on additional MCP connectors.
In an official FreeCAD case, K3.1 Pro built a turbofan engine in Python containing a fan, reduction gearbox, compressor, combustion chamber, turbine and external piping. The output included 2,100 entities, with step-by-step assembly, an internal explosion view and rotational animations for multiple rotors. In another task, the model used trimesh to design a lamp made of five parts that could be 3D printed directly, with design inspiration drawn from Fujian Tulou architecture.
Moonshot shipped this capability directly into the product. The kimi.com website and Kimi App now include a "3D Make" feature. Users can preview and edit models online, download them for 3D printing, or export videos and live photos.
Before the official release, an anonymous model codenamed "Kibitz" had already appeared on Arena. Testers noticed its strong performance in 3D scenes, color selection, water reflection and overall composition. After release, developers began placing K3.1 Pro into real 3D production workflows. One AI game developer used K3.1 Pro with the pi.dev Harness to generate a high-detail "pelican riding a bicycle" 3D scene and concluded that the model, about to open its weights, can already produce quite good 3D results. However, the test also revealed a gap: K3.1 Pro did not reach the final result in a single prompt, requiring multiple rounds of "continue enhancing" instructions. In contrast, Claude Opus 5.5 produced a more complete finished product in one prompt during his earlier testing.
5. Deliberately Reducing Cyber-Offensive Capability Before Weight Release
As code and agent capability improved quickly, Moonshot devoted considerable space to explaining K3.1 Pro's cybersecurity capability and why it deliberately reduced part of its offensive security capability before releasing open weights.
During the expanded post-training phase, the team observed that the model's vulnerability-research ability grew noticeably. Further tracking revealed that some cyber-offensive data had mixed into the mid-training general corpus, and the reinforcement-learning environment also contained corresponding tasks. Internal experiments showed that if this cyber-offensive data was not deleted in either the mid-training or reinforcement-learning stages, the model's downstream offensive capability increased substantially.
When preparing the K3.1 Pro open-weight version, Moonshot identified and deleted the relevant cyber-offensive data from both training stages, while attempting to preserve the general capability required for ordinary programming.
To measure the effect, the team constructed the xdev-bench-hard test. It contains source-code-based vulnerability discovery and exploitation tasks covering recent public vulnerabilities in Linux kernel, Node.js, JavaScriptCore, Exim and VirGL. The agent must read vulnerable source code, prompts and binary files, develop exploits independently, and achieve arbitrary code execution inside an isolated sandbox. If the cyber-offensive training data was not deleted, the same training process produced a model scoring 90.5% pass@3 on xdev-bench-hard. The final released K3.1 Pro drops to 81.0%.
Moonshot's own assessment concludes that K3.1 Pro's cyber-offensive capability still lags frontier closed-source models that have been publicly available for at least three months. This directly influenced the open-weight decision. The principle is twofold: Moonshot does not want an open-weight model to push the frontier of offensive cybersecurity capability; and it believes that, by controlling the speed of capability spread, dual-use models running on user-owned infrastructure can still help more teams audit legacy systems and code lacking professional security review.
Developer opinions diverge on this choice. Some argue that this is equivalent to Moonshot admitting it can internally train a stronger version, simply choosing not to release it. Others worry that reducing the model's security knowledge to lower attack capability might instead harm its ability to discover vulnerabilities and write secure code. A third group views this as a realistic security compromise for a 3-trillion-parameter model about to fully open its weights.
This also surfaces a question rarely faced by open models before: as open-weight models approach the closed-source frontier, "how much capability to open" itself becomes part of model design.
6. Financial Research Gains: Factual Error Rate Drops from 23.4% to 8.3%
Beyond code and 3D, finance is another domain with substantial improvement. On OfficeQA Pro v2, K3.1 Pro scored 64.4, the current best result according to Moonshot's published data. This benchmark is built on real financial documents and tests information retrieval, cross-document verification and multi-step numerical reasoning. On Frontier Finance, which measures a complete investment-research workflow, K3.1 Pro also improved relative to K3.
Moonshot published three internal finance benchmark results. K3.1 Pro improved by 14% over K3 in Financial Modelling, by 20% in Financial Research, and reached roughly 4.2 times K3's score in the Business Operation test that simulates enterprise management and decision-making.
In real cases, K3.1 Pro can generate an Nvidia financial model end-to-end, including business 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 statements, guidance and potential catalysts to form an investment thesis.
For financial agents, the truly hard part is not computation but factual accuracy. Moonshot specifically optimized for factual accuracy in difficult finance and research scenarios. In the internal Kimi Hallu Bench Hard stress test, K3's single-fact statement error rate was 23.4%; K3.1 Pro dropped to 8.3%. The proportion of answers containing at least one factual error fell to about 42.4% of K3's level.
Moonshot emphasized that this is a stress test selecting samples where the model is prone to error. These error rates do not represent overall performance in ordinary usage scenarios.
7. "Quota Burns Out in Minutes" Becomes the Biggest Complaint
In sharp contrast to capability praise, Kimi subscription quotas became the most concentrated negative feedback in the discussion.
One user subscribed to the 99 USD monthly Kimi plan. Running a single session without launching an agent swarm, the user triggered the 5-hour quota limit after roughly 47 minutes. Another user on the 19 USD plan reported that a single task could consume the entire 5-hour quota and visibly erode the weekly quota. One developer stated plainly that no matter how strong the model is, if it exhausts a 5-hour budget in around 30 minutes, it is hard to use it for sustained development.
On Reddit, a more extreme case appeared. A user testing a basic 3D game demo reported that, without sub-agents, the model ran for about 10 minutes, the task had not yet returned a result, and the entire 5-hour quota was already exhausted. Other Plus and higher-tier subscription users reported similar rapid consumption. This led some users to conclude that paying by API for DeepSeek 4.1 Flash is more economical. The stronger the model, the easier it may be to hit the "quota wall".
However, long-horizon agent "quota anxiety" is not unique to Kimi. A Claude user reported that on the 20 USD tier, one complex prompt nearly exhausted the 5-hour quota. Another user, processing an existing codebase with Opus 5.5, triggered the usage limit after about 30 minutes. Many teams rely on cost-effective models such as DeepSeek for large volumes of work, and in the 20 USD plan, code analysis using high reasoning tiers can also exhaust quotas quickly.
Third-party data helps explain much of this quota anxiety. During its full Intelligence Index evaluation, Artificial Analysis recorded that K3.1 Pro cumulatively output about 150 million tokens, while the median model participating in the same evaluation output only 87 million tokens. K3.1 Pro therefore produces roughly 72% more tokens than the median in identical evaluations. Artificial Analysis directly labeled K3.1 Pro as "very verbose".
Speed is not its strength either. Artificial Analysis measured K3.1 Pro's output speed at approximately 73.5 tokens per second, below the 81.1 tokens-per-second median of comparable models. However, many users reported that K3.1 Pro's actual measured speed is far faster than K3.
8. Implications for Engineering Adoption
The token-consumption pattern carries a concrete operational consequence. Teams planning to deploy K3.1 Pro in long-horizon agent pipelines must budget token consumption far more aggressively than for median models. A verbose model that outputs 72% more tokens than peers will multiply both latency and cost in multi-step workflows that repeatedly feed intermediate results back into the context window.
For teams that run K3.1 Pro alongside multiple other model families, centralized access management reduces operational overhead. A unified API gateway abstracts authentication and routing across heterogeneous providers. 4sapi has already synced Kimi K3.1 Pro into its model catalog, enabling one-key multi-use access, so developers can route requests to this new model and other backends through a single credential instead of maintaining separate client configurations for each provider.
9. Conclusion
K3.1 Pro represents a clear step forward in agentic, 3D and financial capability, validated by Terminal-Bench, BenchCAD and internal finance benchmarks. The deliberate reduction of cyber-offensive capability before open-weight release introduces a new design consideration for frontier open models: how much capability should be opened at all. The measured token-consumption pattern, however, turns "very verbose" output into a real operating cost that teams must plan for.
The model is strongest in code, agents, frontend and 3D, while general-knowledge benchmarks improve more slowly. Its open-weight version, with deliberately reduced offensive security capability, still leads ordinary programming capability and is suited to audit work on user infrastructure. For production adoption, teams should combine capability testing with realistic token and quota budgeting, and treat verbose-output behavior as a first-class cost factor rather than a curiosity.
