Moonshot AI launches Kimi K3.1 Pro with 3-trillion parameter scale

A new scale for open weights

Moonshot AI has officially unveiled Kimi K3.1 Pro, marking a significant step forward for the company's model lineup. This new iteration is a 3-trillion-parameter open-weight model, which means the underlying weights of the neural network will be made available to the public. This scale allows the model to handle much more complex reasoning and data processing than smaller alternatives.

The company designed K3.1 Pro as an upgrade to the previous Kimi K3 model. To achieve this, the developers utilized significantly greater computing resources during the post-training phase. These extra resources were intended to refine the model's ability to perform specific, high-level tasks across various industries.

The release is part of a broader strategy to provide high-performance AI tools to a wider audience. By offering an open-weight version of such a large model, Moonshot AI is positioning itself as a major player in the infrastructure for autonomous AI tasks. The full weights are scheduled for release on November 4.

Expanding the context window

One of the defining features of Kimi K3.1 Pro is its 1-million-token context window. A context window determines how much information the AI can "remember" or consider at one time during a single interaction. A million-token capacity allows users to input massive documents, long codebases, or extensive datasets for the model to analyze.

This large capacity is specifically aimed at improving agentic capabilities, which refer to the AI's ability to perform multi-step actions to reach a goal. For example, the model can look at a full project's worth of code to find a bug or analyze a complete financial workflow from start to finish. It is designed to handle these end-to-end deliveries rather than just providing short answers.

To support this, Moonshot AI optimized its reinforcement learning (RL) framework and its long-context processing. These technical improvements ensure that the model remains accurate and efficient even when it is processing very large amounts of information. The goal is to provide a seamless experience for knowledge-based operations that require deep context.

Benchmarking against rivals

Moonshot AI claims that Kimi K3.1 Pro delivers performance that is comparable to leading proprietary models. To prove this, the company cited several specific benchmarks. On Terminal-Bench 4.0, which measures how well coding agents handle real-world terminal-based tasks, the model scored 65.7 percent.

This score is notable because it surpassed the 58.2 percent score of GPT-6 Astra. It also came very close to the 66.4 percent score achieved by Claude Opus 5.5. These comparisons suggest that the model is highly competitive in the space of automated coding and engineering tasks.

The company also highlighted its internal Kimi Code Bench 3.0 Long-horizon. This benchmark tests multi-step execution and tool use across different engineering domains. K3.1 Pro scored 61.2% on this test, which is an improvement over the score of the previous K3 model.

Speed and accessibility

Recognizing that large models can sometimes be slow, Moonshot AI released a high-speed variant called Kimi K3.1 Pro Highspeed. This version is specifically tuned to provide 2-3x the speed of the standard K3.1 Pro model. This makes it a better choice for users who need more responsive interactions and faster turnaround times.

Both the standard and high-speed versions are now integrated into the company's existing ecosystem. Users can access these models through the Kimi.ai website, the Kimi mobile apps, and specialized platforms like Kimi Work and Kimi Code. These tools are designed to cater to different professional needs, from general research to specific coding tasks.

Additionally, the company has made the models available through the Kimi API. An API allows other developers to build their own applications using Moonshot AI's underlying technology. This ensures that the new 3T-class model can be embedded into a wide variety of third-party software and services.

Targeting professional workflows

The rollout of Kimi K3.1 Pro is specifically tailored toward professional and industrial applications. The company highlights its ability to cover the full range of coding work, from making quick fixes in existing codebases to building entire projects from scratch. This versatility is intended to appeal to software engineers who need more than just a simple autocomplete tool.

Beyond coding, the model is positioned to handle complete financial workflows. This implies the ability to process complex data, perform research, and execute a sequence of related financial tasks. By focusing on these high-value areas, Moonshot AI is targeting users who require high-precision output in regulated or data-heavy environments.

The improvements in training efficiency and infrastructure also mean the model is better at autonomous AI tasks. As the company continues to optimize these areas, the model is expected to become more reliable for knowledge-based operations. The upcoming release of the weights in November will be a key moment for the open-source community to begin testing these capabilities.