The paper introduces Kimi K3.1 Pro, a 3T-class open-weight model characterized by a 1-million-token context window and advanced agentic capabilities. It is a post-trained iteration of the Kimi K3 architecture, optimized through significantly scaled reinforcement learning (RL) compute applied across hundreds of downstream domains.

Core Methodology

The model’s performance improvements are driven by a specialized RL infrastructure. The developers optimized the entire technical stack—ranging from long-context inference and training efficiency to sandbox infrastructure and rollout scheduling—to ensure that each unit of RL compute results in predictable, stable gains.

A primary technical focus involves the model’s long-horizon coding and tool-use capabilities. Through large-scale RL, K3.1 Pro achieved proficiency in complex tasks, such as building a human walking dynamics simulator using the *human-atlas* anatomical model. The model demonstrated advanced self-correction logic: when initial reward shaping failed to optimize muscle-moment-arm calculations (a failure in the simulator’s physical dynamics), the model autonomously performed open-loop probes to identify and correct 36 specific mathematical errors before retraining the controller.

In the domain of 3D modeling, K3.1 Pro utilizes CAD code generation to achieve high reconstruction accuracy. On the BenchCAD benchmark, it achieved a 92.7% Voxel Intersection over Union (IoU) score when using tools, calculated as:
$$\text{IoU} = \frac{|A \cap B|}{|A \cup B|}$$
where $A$ represents the ground-truth voxel set and $B$ represents the model-generated reconstruction.

Financial and Factual Performance

K3.1 Pro demonstrates state-of-the-art performance in financial workflows, including multi-step numerical reasoning and cross-document verification. To mitigate common large language model issues, the developers focused on reducing factual hallucination. Using the internal Kimi Hallu Bench, they reduced the claim-level hallucination rate from 23.4% in K3 to 8.3% in K3.1 Pro, a 64.5% relative improvement.

Cybersecurity and Alignment

The paper details a deliberate strategy for managing dual-use cyber capabilities. Rather than relying on simple post-training alignment, the researchers identified and removed cyber-specific data during both the mid-training and RL stages. This "data cleaning" approach reduced the model’s score on the xdev-bench-hard vulnerability discovery task from 90.5% (unrestricted) to 81.0% (final release version). The authors argue that this controlled diffusion is safer than centralizing capabilities in proprietary silos, as it enables developers to audit and secure legacy systems.

Benchmarking and Availability

K3.1 Pro excels in coding and agentic environments, scoring 65.7% on Terminal-Bench 4.0 and 61.2% on Kimi Code Bench 3.0. The model is available across the Kimi ecosystem, with full open-weight model weights scheduled for release by November 4, 2026. The release also includes Kimi K3.1 Pro Highspeed, which provides a 2–3x throughput improvement for low-latency interactions.