Moonshot AI releases Kimi K3.1 Pro with 2.8 trillion parameters

A new standard for Chinese AI

Moonshot AI, a startup based in Beijing, has officially launched its upgraded Kimi K3.1 Pro model. This new iteration is designed to handle complex, multi-step assignments across various professional fields. It specifically targets improvements in 3D design, financial research, and software development.

According to China Daily, the model debuted fourth on the Intelligence Index by independent evaluator Artificial Analysis. It earned a score of 55 points, which represents an 11-point increase over its predecessor. This ranking positions it as a major contender in the global AI landscape.

The company plans to make the model's full weights available to the public by November 4. These weights are the essential files that developers need to run the model on their own hardware. Once published, it is expected to become the highest-ranked open-weight model on the Intelligence Index.

Technical scale and capabilities

The K3.1 Pro is a massive model built with 2.8 trillion parameters. In technical terms, parameters are the internal variables the model uses to learn patterns. For every token the model processes, it activates about 104 billion of those parameters to generate a response.

One of the model's standout features is its one-million-token context window. This allows the system to ingest and process very large volumes of information in a single session. This is particularly useful for analyzing long financial documents or managing extensive coding projects without losing track of the beginning of the conversation.

Moonshot AI achieved these results by further training the original Kimi K3 model. The goal of this specific upgrade was to improve the model's ability to complete practical, real-world tasks. This includes everything from creating 3D assets to performing extended, multi-step assignments that require sustained logic.

Balancing power and security

While the model is highly capable in programming, Moonshot AI has taken specific steps to limit its potential for misuse. The company deliberately removed cybersecurity-related data from parts of its training pipeline. This includes removing data from reinforcement-learning environments, which are used to teach models through trial and error.

The company stated that it does not want to release weights that would extend the absolute frontier of cyber capability. To test this, they conducted internal evaluations on vulnerability discovery and exploitation. The K3.1 Pro scored 81.0 percent on these tests, which was lower than a comparable version that lacked these specific safeguards.

Despite these restrictions, the model remains highly competitive in general programming. In the Terminal-Bench 4.0 test, which measures an agent's ability to perform practical computing tasks, K3.1 Pro scored 65.7 percent. This performance placed it ahead of OpenAI's GPT-6 Astra, which scored 58.2 percent.

Closing the global performance gap

The launch of Kimi K3.1 Pro arrives at a time when the performance gap between Chinese and American AI models is shrinking. A report from Bloomberg Intelligence suggests that the gap has narrowed to around 3 percent, which is the lowest level on record. This trend has led some analysts to suggest that the performance difference has effectively closed.

Experts from Tsinghua University note that Chinese models are now entering the global top tier. They are carving out a unique path by focusing on model efficiency, open-source development, and industrial applications. This shift suggests that Chinese firms are finding ways to compete even as they face challenges in high-end computing chips.

Moonshot AI is also preparing for a significant business expansion. Bloomberg reported that the company recently completed a private funding round that valued it at about $50 billion. The firm is currently targeting a debut on the Hong Kong stock market in the first quarter of 2027.