Moonshot AI launches Kim K3.1 Pro with 3 trillion parameters

A new flagship for complex tasks

Moonshot AI has officially introduced Kim K3.1 Pro, its latest flagship large language model. This new version is a post-trained upgrade of the Kimi K3 model that the Chinese company first released in July. It is designed to compete with major industry models like Claude Opus 5.5 and GPT-6 Astra.

The model is built to handle highly specific and difficult technical workflows. Specifically, it excels in 3D modeling and complex financial analysis. These capabilities make it a targeted tool for engineers and financial professionals who need more than just general text generation.

While it performs well in these specialized areas, its performance varies across other categories. For instance, it significantly trails behind competitors in cybersecurity benchmarks. Moonshot AI stated that engineers intentionally removed cybersecurity data from the training set to manage safety risks.

Efficient architecture at scale

The sheer size of Kim K3.1 Pro is notable, as it contains 3 trillion parameters. To manage this scale, the model is composed of 896 individual neural networks. However, the system is designed to only activate 16 of those networks at any single moment. This approach helps the company minimize inference costs, which are the resources required to generate a response.

Moonshot AI also utilizes a custom technology called Kimi Delta Attention. This is a specific implementation of Gated DeltaNet, which is a memory-saving technique. By using this, the model reduces the amount of RAM needed for its attention mechanism to interpret what a user is asking.

The development of this model required a significant amount of processing power. Moonshot AI noted that the post-training workflow for Kim K3.1 Pro consumed much more power than their previous projects. To handle this, they upgraded their post-training stack and added automated quality checks to improve infrastructure efficiency.

Advancements in 3D and coding

One of the most distinct features of Kim K3.1 Pro is its ability to handle 3D modeling. Rather than drawing objects, the model creates 3D models by writing code snippets that define geometry in numerical terms. It achieved a score of 92.7% on the BenchCAD benchmark, which is a standard test for 3D model coding.

This performance puts the model within four percentage points of its primary competitors. Beyond static models, the AI can also animate objects. In a demonstration, Moonshot AI used the model to create an animated simulation of human motion. The company also used it to generate multiple 3D video games.

Generating games is a difficult task for AI because they require massive amounts of code. Even a single error can cause a game to crash. On the Terminal-Bench 4.0 coding benchmark, the model scored 65.7%. This result put it more than 7% ahead of Astra, though it remained slightly behind Opus 5.5.

Financial analysis and availability

The model also shows high proficiency in corporate and financial environments. It outperformed both Astra and Claude Opus 5.5 on the OfficeQA Pro v2 test. This specific benchmark measures how well a model can perform complex financial analysis tasks, which are often required in professional office settings.

Users can currently access Kim K3.1 Pro through Moonshot AI's paid cloud services. The company is positioning this as a high-end tool for businesses that need reliable performance on complex data. It represents a move toward more specialized, high-utility AI applications.

In a move toward transparency, Moonshot AI has announced a timeline for the model's weights. The company plans to open-source the weights by November 4. This will allow the broader developer community to see the underlying structure of the model once it has been officially launched.

Future safety and features

Moonshot AI is taking a deliberate approach to the safety of its models. By removing cybersecurity-related information during the post-training phase, they aim to reduce the risk of the model being used for malicious purposes. This is a strategic choice to ensure the model remains a safe tool for enterprise use.

However, the company is not abandoning these capabilities entirely for the long term. They have stated that future models will be equipped with more advanced vulnerability analysis features. This suggests a plan to balance safety with utility as the technology matures.

As the model enters the market, it will be a key point of comparison for how 3D and financial tasks are handled by AI. The balance between its massive 3 trillion parameter count and its efficient activation system will be a major point of interest for researchers. It provides a clear look at how companies are trying to scale intelligence while managing costs.