Kimi K3.1 Pro no longer just codes; it builds entire projects.

The Kimi K3.1 Pro combines 2.8 trillion parameters, a 1-million-token context window, and enhanced capabilities in coding, 3D, and finance.

Creating a full video game, building a 3D scene in Blender, developing an application, or producing a financial model requires more than generating code or answering an isolated question. These tasks involve using multiple tools, maintaining context over long sessions, and correcting errors as steps progress. Moonshot AI is developing these capabilities with Kimi K3.1 Pro, an evolution of its K3 multimodal model that claims performance close to that of top proprietary systems across several programming and agentic work benchmarks. The model combines a total of 2.8 trillion parameters, with 104 billion active per token, and a one-million-token context window. Two versions are accessible on Kimi services, while the release of the weights is announced for November 4, 2026. A Large-Scale Model, But with Partial Activation

Kimi K3.1 Pro belongs to the Mixture of Experts (MoE) family of models. Its architecture has 2.8 trillion parameters, but it mobilizes only a portion of them when processing each token. The 104 billion active parameters thus represent the capacity used for a single computation step, without requiring the simultaneous execution of the entire network.

This organization allows for the development of very large models while limiting the amount of computation required for each response. However, it does not make local deployment simple: the full weights must still be stored or distributed across an infrastructure capable of hosting them, and the one-million-token context adds its own memory and compute requirements.

K3.1 Pro retains the multimodal capabilities of its family and can work from documents, images, and textual information. It is a post-training of Kimi K3, and not a new model pre-trained entirely from scratch. Moonshot indicates that it significantly increased the resources dedicated to reinforcement learning, covering several hundred domains and improving the tools necessary for long training runs: execution environments, task management, long-context inference, and evaluation quality control.

The goal is to enable the model to handle projects involving many successive steps, rather than concentrating its performance on short interactions. Complete Projects in Unreal Engine, Unity, and Godot

Programming is one of the main focus areas of this release. Moonshot presents K3.1 Pro as a model capable of developing entire projects using environments such as Unreal Engine, Unity, and Godot, by combining programming, information retrieval, 3D asset generation, and the use of external tools.

Among the demonstrations presented is a celestial palace environment designed in Unreal Engine 5.8, featuring a 36-bay colonnade, a procedurally modeled dragon in Blender, and a character created with MetaHuman. The studio states that the scene uses Nanite to handle its geometry and Lumen for its lighting. The example also relies on several external resources, including materials from Poly Haven and lunar data from NASA.

Moonshot also shows an interactive automotive presentation built with Three.js. The project includes a fabric reveal animation, a virtual garage, and an exploded view of 968 mechanical parts. The cloth simulations and camera movements were prepared in Blender before their integration into the web experience.

These productions demonstrate an ability to coordinate diverse software and resources. They do not mean that every element was created without outside assistance: the demonstrations use libraries, established tools, and certain pre-existing assets. They primarily illustrate the model's ability to drive a development process involving multiple technologies. A Score Close to Claude Opus 5.5 on Terminal-Bench

The programming evaluations published by Moonshot place K3.1 Pro among the top-performing models in the compared selection. On Terminal-Bench 4.0, which measures the ability of agents to accomplish technical tasks in a terminal, the model reaches 65.7%, compared to 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.

The closeness of the scores is not uniformly found across all evaluations. On ProgramBench, K3.1 Pro obtains 31.5%, compared to 50% for GPT-6 Astra and 65% for Claude Opus 5.5. Conversely, on SWE-Marathon v1.1, it reaches 51.7%, ahead of the 46.3% reported for Opus 5.5 in Moonshot's table.

Performance therefore varies depending on the nature of the tasks, the permitted tools, and the scoring methods. Furthermore, the results gather evaluations conducted by Moonshot and figures published by other organizations or vendors, sometimes with different execution environments. This is not a single ranking obtained under perfectly homogeneous conditions.

On its internal Kimi Code Bench 3.0 Long-horizon benchmark, dedicated to long-horizon programming tasks, Moonshot reports an increase from 55.95% for K3 to 61.2% for K3.1 Pro. The lab uses this indicator to measure the effects of its additional training, but the evaluation remains internal to the company. Parametric Modeling Becomes a Core Domain

Beyond programming, K3.1 Pro emphasizes parametric 3D creation. The model can produce scripts for Blender and computer-aided design software, build geometries, and modify an object's parameters without necessarily going through a conversational interface connected to an MCP connector.

This capability applies equally to mechanical parts, architectural elements, configurable objects, or props intended for video games and audiovisual productions. Scripted generation notably allows for maintaining editable geometric relationships, unlike an object created solely as a static mesh.

On BenchCAD, an evaluation requiring the reconstruction of 3D objects from multiple views, K3.1 Pro achieves 92.7% with tool use, compared to 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. Without tools, its score drops to 72.8%, compared to 73% for Opus 5.5. The results with tools cover a subset of 1,000 cases used by Moonshot for its comparisons.

The lab accompanies this evolution with 3D Make, a feature