Unveiling Kimi K3.1 Pro & K3.1 Pro Highspeed

Welcome to the next evolution of open intelligence. We are proud to announce Kimi K3.1 Pro, the flagship of our new K3.1 series. This is our most powerful 3T-class open-weight model to date, featuring a massive 1-million-token context window.

K3.1 Pro introduces significant leaps in:

• 💻 Advanced Coding
• 🎨 3D Asset Creation
• 📈 Professional Financial Research
• 🤖 General Agentic Workflows

For those requiring maximum responsiveness, we are also launching Kimi K3.1 Pro Highspeed, which operates at $\approx 2\text{--}3\times$ the velocity of the standard Pro model.

Availability: Both models are live now via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

---

🚀 Pushing the Boundaries of Open Intelligence

Kimi K3.1 Pro is a post-trained evolution of Kimi K3, fueled by a substantial increase in post-training compute. To ensure this compute translated into actual performance, we overhauled our entire technical stack to support large-scale joint Reinforcement Learning (RL) across hundreds of specialized domains.

The Optimization Stack

To make RL compute stable and predictable, we optimized the following layers:

As a result, K3.1 Pro now rivals or exceeds the capabilities of the leading proprietary models in knowledge, agentic tasks, and programming.

Coding Benchmarks

We tested K3.1 Pro against the industry's best using rigorous, real-world constraints.

---

💻 Comprehensive Coding & Software Engineering

K3.1 Pro isn't just for snippets; it handles everything from minor bug fixes to the architecture of entire projects. It possesses a refined aesthetic sense for web and game development, utilizing the best tools for the job.

Key Integration Capabilities:

• Game Engines: $\text{Unity} \cup \text{Unreal Engine} \cup \text{Godot}$
• Web Tech: Interactive Three.js scenes, polished typography, and fluid animations.
• Workflow: Combines web search with 3D generation for end-to-end product delivery.

Case Study: The Human Walking Simulator

To test long-horizon reasoning, K3.1 Pro was tasked with building a walking dynamics simulator using human-atlas (an open-source model with 296 skeletal and 402 muscle components).

The Iterative Process:

1. Initial Phase: The simulated human collapsed almost immediately.
2. Refinement: Through 167 missions and 420 reviews, the model diagnosed errors and tuned the controller.
3. Breakthrough: When reward shaping failed, K3.1 Pro performed open-loop probes of the dynamics.
4. Correction: It identified and fixed 36 specific errors in shoulder-muscle moment-arm calculations.
5. Result: The simulation achieved continuous alternating strides for 11.2 seconds.

---

🧊 3D Modeling and Spatial Reasoning

K3.1 Pro excels at translating text, sketches, or images into precise 3D geometry. It is particularly adept at parametric modeling for architecture, mechanical engineering, and custom furniture.

Technical Performance (BenchCAD): On the BenchCAD test (3D reconstruction via CAD code), K3.1 Pro achieves:

• With Tools: $92.7\%$ (Close to Claude Opus 5.5's $96.2\%$ and GPT-6 Astra's $95.9\%$)
• Without Tools: $72.8\%$ (Outperforming most competitors)

Tooling & Workflow: The model can write and execute scripts for Blender and various CAD software to manipulate:

• $\text{Geometry} \rightarrow \text{Materials} \rightarrow \text{Scene Layout}$

New Feature: 3D Make Now available on the Kimi app and website, 3D Make allows you to:

• Preview and edit 3D models in real-time.
• Download files optimized for 3D printing.
• Export as Live Photos or videos.

---

💰 Professional Finance & Analytical Rigor

Financial workflows require absolute precision. K3.1 Pro provides end-to-end delivery, from initial screening to final profitability analysis.

Performance Metrics

• OfficeQA Pro v2: Scored 64.4 (State-of-the-Art), excelling in cross-document verification and numerical reasoning.
• Frontier Finance: Significant improvement over K3 in investment workflow support.

Internal Finance Benchmarks

We measure success across three critical dimensions:

1. Financial Modeling: Accuracy of formula linkages and workbook logic.
2. Financial Research: Ability to produce auditable research within strict cutoff constraints.
3. Business Operation: Decision-making within simulated corporate environments.

Eliminating Hallucinations

To ensure reliability, we utilized the Kimi Hallu Bench Hard stress test. The results show a dramatic reduction in factual errors:

$\text{Claim-level Hallucination Rate: } 23.4\% \text{ (K3)} \longrightarrow 8.3\% \text{ (K3.1 Pro)}$

High error rates $\rightarrow$ High factual integrity. The share of responses containing at least one error is now only 42.4% of what it was in the previous version.