Moonshot AI Releases Kimi K3.1 Pro Multimodal MoE Model

Introduction

On October 23, 2026, Moonshot AI marked a historic milestone in the global AI landscape with the release of Kimi K3.1 Pro. As the flagship of the new K3.1 series, this model represents a massive leap in scaling laws and post-training efficiency, positioning Moonshot AI at the absolute forefront of the 3T-class model race.

For developers and AI engineers, Kimi K3.1 Pro is more than just a larger parameter count; it is a sophisticated multimodal agent capable of navigating complex, long-horizon tasks that previously required human-in-the-loop intervention. Whether it is simulating complex biological dynamics or managing full-stack game development, K3.1 Pro is designed for the next generation of autonomous workflows.

• First in the K3.1 series
• Native multimodality with 1M token context window
• Designed for high-stakes engineering and financial reasoning

Key Features & Architecture

The architecture of Kimi K3.1 Pro is built on a massive Mixture-of-Experts (MoE) framework. It boasts a total of 2.8 trillion parameters, with 104 billion active parameters per token. This configuration allows the model to maintain the reasoning depth of a massive dense model while retaining the inference efficiency required for real-world deployment.

A critical differentiator is the training methodology. K3.1 Pro is a post-trained evolution of the original Kimi K3, benefiting from significantly larger post-training compute. Moonshot AI optimized every layer of the stack—from the RL framework and long-context inference engines to the sandbox infrastructure and rollout scheduling—to ensure stable, predictable returns for every increment of reinforcement learning compute.

• Total Parameters: 2.8T
• Active Parameters: 104B (MoE)
• Context Window: 1,000,000 tokens
• Optimized Stack: Custom RL framework and sandbox infrastructure

Performance & Benchmarks

Kimi K3.1 Pro delivers state-of-the-art results across critical agentic and coding benchmarks. In Terminal-Bench 4.0, which tests resource-constrained terminal tasks with strict validation, the model scored 65.7%. This places it ahead of GPT-6 Astra (58.2%) and brings it within striking distance of Claude Opus 5.5 (66.4%).

Coding capabilities have seen a massive jump via the Kimi Code Bench 3.0 Long-horizon. The model achieved 61.2%, a significant improvement over the previous K3's 55.95%. This improvement is driven by the model's ability to handle multi-step execution, tool use, and end-to-end problem solving across diverse engineering domains, including full-project development in Unity, Unreal Engine, and Godot.