Moonshot AI launches Kimi K3.1 Pro to challenge US coding models

A new frontier for Chinese AI

Moonshot AI, a startup based in Beijing, has officially introduced Kimi K3.1 Pro. This model serves as the debut for the K3.1 series and represents the company's most capable open-weight offering in the three-trillion-parameter class to date.

The new system was built by significantly increasing post-training computation compared to the previous Kimi K3. The developers used large-scale reinforcement learning across hundreds of different domains to refine the model's capabilities.

As reported by Crowdfund Insider, the model is designed to compete directly with frontier systems from the United States. It aims to prove that high-performing AI can be developed outside of the most dominant Western tech hubs.

Architecture and technical specs

Kimi K3.1 Pro utilizes a mixture-of-experts architecture. This means the model has about 2.8 trillion total parameters, but only roughly 104 billion active parameters per token, which helps manage the computational load during use.

The model also features a one-million-token context window, allowing it to process and remember very large amounts of information at once. It supports multiple types of inputs, including text, images, and in some descriptions, video content.

For users who need speed, Moonshot AI offers a Highspeed variant. This version is designed to generate responses two to three times faster than the standard Kimi K3.1 Pro while maintaining its core capabilities.

Coding and 3D design power

One of the primary focuses of this update is stronger coding and agentic performance. In tests on Terminal-Bench 4.0, the model scored 65.7 percent on tasks involving realistic, resource-constrained terminal operations with strict verification.

This score reportedly places Kimi K3.1 Pro ahead of OpenAI’s GPT-6 Astra. It also demonstrated the ability to handle full-project work in popular game engines like Unity, Unreal, and Godot, as well as creating interactive web experiences.

The model also shows high aptitude in three-dimensional design. On BenchCAD, it scored 92.7 percent when using tools to reconstruct 3D objects from images. It can drive software like Blender to edit geometry, materials, and scenes for mechanical and architectural parts.

Accuracy and safety measures

Moonshot AI targeted factual accuracy to reduce the rate of hallucinations, which are instances where an AI generates false information. They used an internal hard benchmark to measure these improvements over the previous generation.

The company also highlighted success in financial workflows. The model showed state-of-the-art results in multi-step numerical reasoning over real documents and simulated business-operation tasks.

To prepare for the open-weight release, the company applied specific data-filtering measures. These filters were intended to restrain certain vulnerability-research capabilities, trading some specialized performance for a reduction in potential misuse risk.

The road to open weights

While the models are live now on Kimi’s website, mobile apps, and API, the full weights are not yet public. The company has promised to release the full model weights to the public by November 4, 2026.

Early independent evaluations place the model fourth on the Artificial Analysis Intelligence Index. Once the weights are released, it is expected to become the highest-scoring openly available model on that specific index.

The release of Kimi K3.1 Pro raises significant questions about the durability of U.S. leadership in frontier AI. It suggests that the performance gap between Chinese open-weight offerings and top proprietary systems may be closing rapidly.

Comparing the competition

In the company's published comparisons, the score on Terminal-Bench 4.0 places the model ahead of OpenAI’s GPT-6 Astra. It sits only narrowly behind Anthropic’s Claude Opus 5.5, showing how it stacks up against the most prominent systems from the United States.

This follows the momentum of the previous Kimi K3 release from this past summer. That earlier model already garnered significant attention by posting competitive results that raised questions about the durability of U.S. leadership in the frontier artificial intelligence space.

Advanced simulation and CAD

The model demonstrated its ability to handle complex, long-horizon coding tasks in internal evaluations, showing clear gains over its predecessor. In one specific demonstration, it iteratively improved a simulator for human-walking dynamics built on an open anatomical model. Instead of just adjusting rewards, it diagnosed specific issues and corrected underlying simulation errors.

The system also shows high proficiency in reconstruction tasks on BenchCAD. By generating CAD code from multi-view images, it achieved a score that approaches the results of leading proprietary systems. This capability is particularly useful for creating parametric models of architectural forms and configurable products.