Moonshot AI's 2.8T-parameter vision-language model for long-horizon coding, 3D creation, and financial research, with 104B active parameters and a 1M-token context window.
Kimi K3.1 Pro is an open-weight, natively multimodal model from Moonshot AI. Post-trained from Kimi K3 with large-scale reinforcement learning across hundreds of domains, it improves coding, tool use, spatial reasoning, and document-based research. It accepts text and images and produces text.
K3.1 Pro is built for multi-step engineering work: navigating codebases, debugging, using tools, and carrying projects through implementation and testing. Moonshot AI also highlights frontend development and work with Unity, Unreal Engine, and Godot.
Its spatial reasoning supports generating CAD code and scripts for tools such as Blender from text briefs or reference images. These workflows rely on external tools to build and render assets; the model itself outputs text.
K3.1 Pro combines visual document understanding with retrieval, cross-document verification, and multi-step numerical reasoning. Moonshot AI reports improvements over K3 in financial modeling and source-grounded research, alongside a reduction in factual errors on its internal hallucination benchmarks. These results do not eliminate the need to verify sources and calculations.
Results are reported by Moonshot AI in the official model card and technical blog.
| Benchmark | Kimi K3.1 Pro |
|---|---|
| Terminal-Bench 4.0 | 65.7 |
| DeepSWE v1.1 | 75.5 |
| SWE-Marathon v1.1 | 51.7 |
| BenchCAD, with tools (Voxel IoU) | 92.7 |
| OfficeQA Pro v2 | 64.4 |
| MMMU-Pro, without tools | 84.1 |
Reported evaluations use max reasoning effort, temperature 1.0, and top-p 0.95 unless otherwise specified. Terminal-Bench, DeepSWE, and SWE-Marathon use the Claude Code harness. BenchCAD uses a 1,000-case subset and Claude Code for tool-assisted evaluation. OfficeQA provides the PDF corpus as page images without machine-readable text. Scores use each benchmark's own metric and are not directly comparable across rows. See the sources for full methodology.
K3.1 Pro combines Kimi Delta Attention with Gated DeepSeek Sparse Attention across 93 layers. Its latent MoE selects 16 of 896 routed experts per token, alongside two shared experts. MoonViT-V2 encodes images, and quantization-aware training uses MXFP4 weights and MXFP8 activations.
The open weights are released under the Kimi K3.1 Pro License.