Kimi K3 1 Pro Open Weight Modell Image © Moonshot AIKimi K3 1 Pro Open Weight Modell (Image © Moonshot AI)

In addition to the standard version, the company has released “Kimi K3.1 Pro Highspeed,” a variant optimized for lower latency that delivers responses two to three times faster than the base model.

Kimi K3.1 Pro is a fine-tuned version of the previous Kimi K3 model. The development process involved a significant increase in computing resources for retraining, as well as a comprehensive overhaul of the training infrastructure.

Kimi K3 1 Pro Open Weight Modell HallucinationKimi K3 1 Pro Open Weight Modell Hallucination (Image © Moonshot AI)

Moonshot AI has optimized several layers of the technical stack to support large-scale collaborative reinforcement learning (RL) across hundreds of domains. These optimizations included the RL framework, training efficiency, the sandbox infrastructure, rollout planning, and inference with long contexts. These changes were implemented to ensure that the increase in RL computing power leads to stable and predictable performance gains across all tested domains.

Kimi K3 1 Pro Open Weight Modell PreviewKimi K3 1 Pro Open Weight Modell Preview (Image © Moonshot AI)

Software Development and 3D Modeling Performance

The model is designed to handle a wide range of programming tasks, ranging from small bug fixes to the creation of complete software projects. It is compatible with major game engines, including Unity, Unreal Engine, and Godot, and can integrate web search into 3D generation tools.

Kimi K3 1 Pro Open Weight Modell Terminal BenchKimi K3 1 Pro Open Weight Modell Terminal Bench (Image © Moonshot AI)

In standardized tests, Kimi K3.1 Pro achieved a score of 65.7% on Terminal Bench 4.0, which evaluates programming agents on resource-constrained terminal tasks. This result is higher than the 58.2% achieved by GPT-6 Astra and slightly lower than the 66.4% achieved by Claude Opus 5.5. Additionally, the model scored 61.2% on the internal “Kimi Code Bench 3.0 Long-horizon” test—an improvement over the 55.95% achieved by Kimi K3.

In the area of 3D modeling and spatial reasoning, the model achieved a score of 92.7% when using tools on BenchCAD. By comparison, Claude Opus 5.5 and GPT-6 Astra scored 96.2% and 95.9%, respectively. Without the use of tools, K3.1 Pro scored 72.8%, placing it just behind Claude Opus 5.5 at 73%.

Cybersecurity Controls and Safety Measures

As the model’s general capabilities increased, Moonshot AI implemented targeted controls to mitigate the risk of misuse in offensive cybersecurity operations. The company identified and removed cyber-related data from both the training corpus and the reinforcement learning environment.

The effectiveness of these controls was measured using xdev-bench-hard, a benchmark focused on detecting and exploiting vulnerabilities. A version of the model without these security measures solved 90.5% of the tasks on pass@3. The final version, K3.1 Pro—with the data removed—solved 81.0%. Moonshot AI explained that this approach allows the model to remain useful for defensive security audits and the maintenance of legacy systems without expanding the absolute limits of offensive cyber capabilities.

Availability and Access

Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are currently available through the following channels:

  • Web and Mobile: Kimi.ai and the Kimi apps for iOS, Android, and HarmonyOS.
  • Professional Tools: Kimi Work (desktop version 3.1.0 or higher) and Kimi Code.
  • API: The Kimi API platform with a “Zero Data Retention” option.

The full model weights are scheduled to be released by November 4, 2026. Company administrators can enable the model for their organizations via the Kimi Business settings.