Moonshot AI releases Kimi K3.1 Pro, 2.8 trillion parameters, coding agent benchmark surpasses GPT-6 Astra

Chinese large-model company Moonshot AI has launched its new-generation flagship model Kimi K3.1 Pro, targeting complex professional tasks such as programming, 3D creation, and financial research, and plans to release the full model weights by November 4. The model has a total parameter scale of 2.8 trillion, adopts a mixture-of-experts (MoE) architecture, activates 104 billion parameters at a time, and supports a 1-million-token context window.

Moonshot AI says Kimi K3.1 Pro is the first model in the K3 series and the company's most capable 3-trillion-level parameter model to date. Compared with earlier models that focused on general capabilities, the new model emphasizes long-horizon programming, 3D modeling, financial analysis, and the ability of agents to execute complex tasks, and reaches or approaches the level of top closed-source models on multiple tests.

Moonshot AI also launched a high-speed version, Kimi K3.1 Pro Highspeed, saying its intelligence level is comparable to the standard version while output speed can increase to about 2 to 3 times. Currently, both models are available through the Kimi official website, App, Kimi Work, Kimi Code, and API, and the full model weights are expected to be released by November 4.

For Moonshot AI, the core of the Kimi K3.1 Pro upgrade is not just pushing the parameter scale to the 3-trillion level, but further integrating programming, spatial reasoning, financial analysis, and long-horizon agent execution capabilities into a single model.

As the full weights are about to be opened, what the market needs to watch next is whether these benchmark advantages can be consistently delivered in real projects, and whether developers can turn the model's capabilities into stable, reliable productivity at an acceptable deployment cost.

Backed by large-scale reinforcement learning, complex task execution improves

Kimi K3.1 Pro is not a new model trained from scratch, but a post-trained version based on Kimi K3. Moonshot AI says the new model substantially increased post-training compute and optimized areas including the reinforcement learning framework, long-context reasoning, training efficiency, sandbox infrastructure, and task scheduling systems to support large-scale joint training across hundreds of downstream domains.

Moonshot AI believes the key to this training route is converting larger-scale reinforcement learning investment into sustained execution capability in complex tasks, rather than merely improving single-turn question-answering performance.

Moonshot AI says that on code, Agent, and knowledge benchmarks, K3.1 Pro reaches or approaches the level of the strongest closed-source models. It disclosed that on the programming agent test Terminal-Bench 4.0, Kimi K3.1 Pro scored 65.7%, higher than OpenAI's GPT-6 Astra at 58.2%, and only slightly behind Anthropic's Claude Opus 5.5 at 66.4%.

Terminal-Bench 4.0 tests the model's ability to use tools, execute commands, and solve programming problems under resource-constrained conditions through terminal tasks in real environments.

On Moonshot AI's internal Kimi Code Bench 3.0 Long-horizon test, the new model scored 61.2%, higher than the previous generation K3's 55.95%. This test focuses on multi-step execution, tool calling, and end-to-end engineering problem-solving capability.

It should be noted that the above results come from different testing systems, some of which are Moonshot AI internal benchmarks. Test environments, tool configurations, and evaluation methods may differ across models, and scores cannot be directly equated with capability rankings in all real work scenarios.

From writing code to developing games, attempting to complete entire projects end to end

In programming, Moonshot AI hopes Kimi K3.1 Pro can handle tasks ranging from fixing problems in existing codebases to developing complete software projects from scratch.

Cases demonstrated by Moonshot AI include developing games using game engines such as Unreal Engine, Unity, and Godot, as well as generating web pages, Three.js interactive scenes, and 3D assets.

Among them, in Unreal Engine 5.8, the model built an Eastern mythological heavenly palace above a sea of clouds. Moonshot AI says the colonnade in the project was procedurally modeled through Blender, the source model contains 36 million triangles, and it was rendered using Nanite and Lumen.

Another case is a 3D action-adventure game developed based on Godot, including five explorable regions, procedurally generated terrain and villages, third-person combat, a two-stage boss fight, a day-night cycle, and fast travel.

The model also completed a browser-based 3D robotic arm display page, generating 2,317 parts and 268 dynamic cable routes, and implementing exploded views, transparent shells, and cross-section displays.

These cases attempt to show that the model's goal is no longer limited to generating code snippets, but further taking on project planning, invoking professional tools, generating assets, debugging, and iteration.

Moonshot AI also demonstrated a more complex simulation task: the model built a human walking dynamics simulator based on an open-source human anatomy model, and continuously adjusted the walking controller across 167 tasks and 420 independent review processes, ultimately enabling the simulated human body to walk continuously for 11.2 seconds.

More notably, when adjusting the reward mechanism failed to improve the arm's movement, the model turned to examining the simulator dynamics, located and corrected 36 errors in the calculation of the shoulder muscle moment arms, and then retrained the controller.

Such cases reflect an important direction for long-horizon agent capabilities: the model must not only generate solutions, but also locate problems based on experimental feedback and continuously improve results across multiple rounds of execution. However, the demonstration cases themselves cannot represent that the model can stably reproduce the same performance in all similar tasks.

Enhanced 3D modeling capabilities, with support for professional tools such as Blender and CAD

Beyond programming, 3D creation and spatial reasoning are also key focuses of this Kimi K3.1 Pro upgrade.

Moonshot AI states that Kimi K3.1 Pro excels at 3D creation and spatial reasoning, and is skilled at parametric modeling, such as mechanical parts, architectural models, multi-parameter configurable products, and custom furniture. On the BenchCAD benchmark, the model achieved a score of 92.7% with tools, close to Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%.

BenchCAD mainly assesses whether a model can reconstruct a 3D model by generating CAD code based on rendered images from multiple viewpoints. The model needs to understand the object's shape, proportions, and spatial relationships, and then convert visual information into a usable three-dimensional structure.

Moonshot AI states that Kimi K3.1 Pro can infer object structures from text descriptions, sketches, or reference images, and directly write and run scripts for professional tools such as Blender and CAD to complete parametric modeling, geometric adjustments, material modifications, and scene layout, without relying on MCP connectors.

The cases demonstrated by Moonshot AI include building an airport indoor rainforest scene in Blender, using FreeCAD and Python to generate a turbofan engine model containing 2,100 entities, and designing a desk lamp that can be directly 3D printed.

Among them, the turbofan engine model supports step-by-step assembly, internal structure display, and rotation animations of multiple rotors; the desk lamp draws on the circular architectural form of Fujian tulou and is designed as a hollow structure composed of five parts.

At the same time, the Kimi official website and App have launched the 3D Make feature, which supports online preview and editing of 3D models, and allows users to download models for 3D printing, or export videos and Live Photos.

This means Moonshot AI is attempting to further transform the model's spatial reasoning and code generation capabilities into directly usable creative tools.

Financial research becomes a key breakthrough area, with the factual error rate declining

Financial research is another key strengthened capability of Kimi K3.1 Pro.

Moonshot AI states that the new model can complete relatively full workflows around financial modeling, investment analysis, and industry research, rather than merely answering financial questions or summarizing research reports.

On the OfficeQA Pro v2 benchmark, Kimi K3.1 Pro scored 64.4 points, and Moonshot AI states that it reaches the current state of the art (SoTA). This test focuses on complex data analysis based on real financial documents, involving information retrieval, cross-document verification, and multi-step numerical reasoning.

In the company's internal financial modeling test, the new model improved by 14% over K3; in the financial research test it improved by 20%; and in the business operations test its score reached about 4.2 times that of K3.

In practical application cases, the model generated a financial model for Nvidia, supporting adjustments to operating assumptions, formula linkages, and sensitivity analysis; it also produced an investment committee report for CATL, sorting out the differences between market consensus expectations and the company's actual performance, and organizing possible future catalysts.

Moonshot AI also demonstrated a quantum computing industry research report, emphasizing that the data and citations in the report can all be traced to primary sources.

However, financial research has far higher accuracy requirements than general content generation. The fact that the model can produce a complete financial workbook or research report does not mean that every piece of data, assumption, and conclusion in it is necessarily correct; actual use still requires verification of key figures and their sources.

To address this issue, Moonshot AI specifically conducted a factual accuracy stress test.

In the internal Kimi Hallu Bench Hard test, the new model's error rate on individual factual statements dropped from 23.4% in K3 to 8.3%. The proportion of responses containing at least one factual error dropped to about 42.4% of the previous generation.

Moonshot AI specifically points out that this test deliberately selects high-difficulty scenarios in which models are prone to errors, and the test results do not represent the overall error rate in daily use.

Nevertheless, reducing the factual error rate remains an important condition for models to enter professional scenarios such as finance and enterprise operations. For these tasks, whether the model can provide traceable evidence, verify calculation processes, and promptly detect errors may be more important than simply increasing response speed.

Standard edition API output priced at 68 yuan per million tokens; safety evaluation before opening weights

On the commercialization side, Moonshot AI has adopted an arrangement in which product services come first and model weights are opened later.

Currently, Kimi K3.1 Pro has been integrated into the Kimi official website, App, Kimi Work, Kimi Code, and Kimi API. Developers can call the model through the API, and enterprise users can use it for programming, research, and other agent workflows.

According to the pricing published by Moonshot AI, the API input price for the standard edition of Kimi K3.1 Pro is 1.3 yuan per million tokens for cache hits and 13 yuan for cache misses, with an output price of 68 yuan. The high-speed edition's output speed is about 2 to 3 times that of the standard edition, at twice the standard edition's price.

Moonshot AI also launched the Kimi Hosted Agents managed agent service, providing enterprises and developers with a cloud-based runtime environment that supports long-duration task execution and multi-agent collaboration, reducing the cost of building and maintaining agent runtime frameworks on their own.

What has drawn more attention from the open-source community is the release plan for the model weights. Moonshot AI said the full weights will be opened up before November 4. This will give developers the opportunity to deploy, research, and adapt the model on their own infrastructure, though the specific scope of use will still be subject to the final released license terms.

At the same time, Moonshot AI also disclosed a cybersecurity evaluation of the open-weight model. The company said that while preparing to open the Kimi K3.1 Pro weights, it had identified and removed cyber offense and defense-related data from the mid-training and reinforcement learning stages to limit the model's vulnerability research capabilities.

In its in-house xdev-bench-hard test, the model without the relevant data removal processing scored 90.5%, while the final version scored 81.0%. The company said this shows that targeted data cleaning reduced the model's vulnerability research capabilities, but stronger general programming and long-horizon debugging capabilities may still transfer to cybersecurity tasks.

This also reflects the dilemma facing open-weight models: on the one hand, open models help enterprises and researchers deploy locally, inspect code, and conduct defensive security work; on the other hand, the model's growing capabilities also require developers to assess its potential risks in dual-use scenarios such as cyber offense and defense. Risk disclosure and disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this article is at the user's own risk.