Kimi K3.1 Pro: Moonshot AI releases weights for its new multimodal model

Moonshot AI has officially released the open weights for Kimi K3.1 Pro, its most capable 3T-class model to date. Built as a natively multimodal agentic system, it is the first model in the K3.1 series, post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of downstream domains.

• Kimi K3.1 Pro open weights are now available for developers and researchers on Hugging Face.
• Features a 1-million-token context window powered by Gated DeepSeek Sparse Attention layers.
• Delivers major performance upgrades in coding, 3D creation, financial research, and agentic workflows.

Architecture and Context Window

A key architectural upgrade in Kimi K3.1 Pro is the replacement of full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers. This modification optimizes long-context modeling, allowing the model to effectively process up to 1 million tokens without sacrificing performance.

The model's evaluation utilized an extensive suite of benchmarks, including Terminal-Bench 4.0, SWE-Marathon, ProgramBench, and BenchCAD. Most evaluations were conducted using maximum reasoning effort settings alongside testing harnesses such as Claude Code to ensure standardized and rigorous scoring.

Multimodal and Agentic Capabilities

Kimi K3.1 Pro is designed from the ground up as a multimodal agent capable of operating across diverse digital environments. It underwent rigorous testing on specialized agentic benchmarks such as OSWorld, OfficeQA Pro, Frontier Finance, and Harvey Legal Agent.

With integrated tool use and advanced visual perception—evaluated via benchmarks like MMMU-Pro and PerceptionBench—the model excels in tasks ranging from graphic design and document analysis to complex software engineering and long-horizon programming.

Impact on the AI Ecosystem

The release of Kimi K3.1 Pro open weights provides the global developer community with a powerful open alternative for building sophisticated autonomous agents. By combining a massive context window with advanced reasoning and multimodal tools, Moonshot AI continues to drive innovation in the open-weight AI landscape.