Moonshot AI launches Kimi K3.1 Pro with 1 million token context window

A new standard for open weights

Moonshot AI, a Chinese artificial intelligence company, has officially launched Kimi K3.1 Pro. This new model is part of the company's K3.1 series and represents their most capable 3T-class open-weight model to date. Open-weight models are those where the underlying parameters are shared with the public, allowing for broader customization and research.

The company developed Kimi K3.1 Pro as an upgraded version of its previous Kimi K3 model. To achieve this, the team utilized significantly greater computing resources during the post-training phase. This investment in infrastructure and training efficiency was intended to improve performance across hundreds of different application areas.

The release is notable because of the company's commitment to transparency. While the model is available now on various platforms, Moonshot AI has scheduled the release of the full model weights for November 4. This move allows the developer community to access the core architecture of a high-performing 3-trillion-parameter system.

Handling massive amounts of data

One of the standout technical features of Kimi K3.1 Pro is its 1-million-token context window. A token is a basic unit of text, and a million-token window means the model can "remember" and process a very large amount of information in a single session. This is particularly useful for tasks that require looking at long documents or large files simultaneously.

The company optimized several key frameworks to make this possible. They focused on long-context processing and reinforcement learning (RL), which is a method of training models through trial and error to improve decision-making. These optimizations were designed to enhance the model's ability to handle autonomous AI tasks and knowledge-based operations.

By improving these underlying systems, Moonshot AI aims to provide a more stable experience for complex queries. The model is built to handle diverse engineering domains, focusing on multi-step execution and tool use. This allows the AI to solve problems that require several sequential steps rather than a single, immediate answer.

Performance in coding and finance

Moonshot AI claims that Kimi K3.1 Pro delivers substantial improvements in specific professional fields. It is designed to cover the full range of coding work, from making quick fixes in existing codebases to building entire projects from scratch. The model is also built to handle end-to-end delivery of complete financial workflows, which often involve complex data analysis.

To prove its capabilities, the company cited several benchmarks. On Terminal-Bench 4.0, which evaluates how well coding agents perform real-world terminal-based tasks, Kimi K3.1 Pro scored 65.7 per cent. This score surpassed GPT-6 Astra's 58.2 per cent and came very close to the performance of Claude Opus 5.5, which scored 66.4 per cent.

The company also used an internal metric called Kimi Code Bench 3.0 Long-horizon. This benchmark tests how well an AI agent can solve engineering problems using tools and multi-step execution. In this test, Kimi K3.1 Pro scored 61.2%, which is an improvement over the previous K3 model.

Speed and accessibility

Recognizing that speed is a priority for many users, Moonshot AI also released a high-speed variant called Kimi K3.1 Pro Highspeed. This version is designed to provide more responsive interactions for people who need faster results. According to the company statement, the high-speed variant delivers 2-3x the speed of the standard Kimi K3.1 Pro model.

Users can access these models through several different channels provided by the company. The models are currently available on the Kimi.ai website, as well as through the Kimi mobile apps. They are also integrated into specialized platforms like Kimi Work and Kimi Code, which are tailored for specific professional tasks.

Additionally, the company has made these models available through the Kimi API. An API, or Application Programming Interface, allows other software developers to integrate the Kimi models into their own applications. This ensures that the new 3T-class model can be used by a wide variety of third-party tools and services across the internet.

What comes next for the series

The launch of Kimi K3.1 Pro marks the beginning of a new series for Moonshot AI. By releasing the weights in November, the company is positioning itself as a major player in the open-weight space. This allows them to compete with leading proprietary models by offering a high-performance alternative that developers can run and modify themselves.

The focus on agentic capabilities suggests that Moonshot AI wants Kimi to do more than just answer questions. They want the model to act as an agent that can perform tasks, such as completing financial workflows or building software projects. This shift toward autonomous tasks is a key trend in the current AI landscape.

As the community begins to experiment with the weights in November, the focus will likely shift to how the model handles specific niche tasks. Developers will be looking to see if the 1-million-token window holds up under extreme stress tests. The success of this release will determine how Moonshot AI scales its K3.1 series for future updates.