Kimi K3.1 Pro distribution extends to independent infrastructure

Moonshot AI releases the weights for Kimi K3.1 Pro, a multimodal model with a one-million-token context window, designed for coding and agentic tasks.

Publishing an artificial intelligence model does not necessarily mean that its users can run it on their own infrastructure. API-accessible services generally keep their internal parameters on their publisher’s servers. Moonshot AI is taking a different approach with Kimi K3.1 Pro: after its commercial launch in October, the Chinese company released the model weights on Hugging Face, in line with its announced schedule.

This release concerns a particularly large multimodal system, with 2.8 trillion parameters and a context window reaching 1,048,576 tokens. The model is designed to process text, interpret images, and carry out complex tasks involving multiple tools. Its release makes it possible to deploy and adapt it independently of the official API, provided suitable computing resources are available and its license is respected. A giant architecture, but only partially activated

Kimi K3.1 Pro’s size is based on a Mixture of Experts, or MoE, architecture. Unlike a dense model, which uses all of its parameters during each computation, this organization selects only part of its components to process each token.

Of the 2.8 trillion parameters announced, about 104 billion are activated at each step. The architecture includes 896 specialized experts, 16 of which are selected per token, plus two shared experts. The network has 93 layers.

This distribution reduces the amount of computation needed at each step compared with activating all the parameters, without eliminating the challenge of storing and operating a model of this size. The released files take up about 1.56 TB in their quantized representation, a volume that places full deployment beyond the reach of ordinary personal computers.

To improve long-context processing, the teams replaced part of Kimi K3’s attention mechanism with a variant based on Gated DeepSeek Sparse Attention. The technical goal is to limit the operations required when the model works on very large amounts of information.

The visual encoding side relies on MoonViT-V2, a 401 million-parameter component. The model can therefore combine image analysis with text instructions, but its published configuration does not correspond to a native image or video generator. Results close to proprietary models on several tasks

Moonshot AI’s technical report presents comparisons with GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1, and several Chinese models.

On Terminal-Bench 4.0, an evaluation of computer tasks performed in a terminal, Kimi K3.1 Pro reaches 65.7%, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5 in the tested configurations. On SWE Marathon v1.1, which examines the ability to solve programming tasks requiring multiple steps, its score reaches 51.7%, compared with 46.3% for Opus 5.5 and 42.5% for GPT-6 Astra.

The results are not, however, uniformly favorable. On Terminal-Bench-Science 0.1, Kimi scores 38.1%, while GPT-6 Astra reaches 68.1%. On Program Bench, the Chinese model records 31.5%, compared with 65% for Opus 5.5.

These gaps show that a model’s standing depends heavily on the tasks and environments used. The tables provided by Moonshot serve as a basis for comparison, but they do not demonstrate overall superiority over proprietary models.

The evaluations were mainly carried out with the highest reasoning level. For several competitors, the configurations also mention fallback mechanisms. The conditions are therefore not necessarily identical across all solutions. From programming to three-dimensional work

Interactive content creation and 3D modeling occupy an important place in the model’s technical presentation.

On BenchCAD, an evaluation covering three-dimensional reconstruction and computer-aided design code generation, Kimi K3.1 Pro reaches 92.7% with tools enabled. Claude Opus 5.5 scores 96.2% and GPT-6 Astra 95.9% in the same comparison.

The model can use text descriptions, sketches, and visual references to produce scripts intended for software such as Blender. The demonstrations include the creation of parametric objects, interactive scenes, and environments for game engines such as Unity, Unreal Engine, and Godot.

A longer internal experiment focuses on human walking simulation. Working on an anatomical representation comprising 296 skeletal elements and 402 muscle components, the model reportedly completed 167 missions and 420 successive revisions. It allegedly identified 36 errors in the calculation of shoulder movements before continuing to optimize the simulator.

These achievements describe capabilities obtained in instrumented environments, with tools and correction loops. They do not guarantee that a single instruction will systematically produce a usable project without human intervention.

Financial applications are also among the fields highlighted. On OfficeQA Pro v2, devoted to working with financial documents, the model reaches 64.4%. On an internal test designed to measure difficult factual errors, Moonshot reports a reduction in the rate of incorrect statements from 23.4% with Kimi K3 to 8.3% with K3.1 Pro. This comparison remains specific to the company’s protocol. Weights available under a specific license

The parameters are available in the official Hugging Face repository, along with a technical sheet and recommendations for vLLM, SGLang, and TokenSpeed environments.

Their release does not amount to an unconditional availability. The Kimi K3.1 Pro license notably allows the use, modification, deployment, adaptation, and redistribution of the components, including in commercial projects.

However, it introduces particular obligations for hosted model providers. A company directly offering inference or model adaptation capabilities to third parties must enter into a separate agreement with Moonshot AI if its consolidated revenue, together with that of its affiliates, exceeds 20 million dollars over twelve consecutive months.

Commercial products exceeding 100 million monthly active users or 20 million dollars in monthly revenue must