Kimi K3.1 Pro Released as Open-Weight Agentic Multimodal Model
New 2.8T parameter model features one-million-token context window and advanced reasoning capabilities.
1New Agentic Model Launch
Moonshot AI has introduced Kimi K3.1 Pro, an open-weight, natively multimodal agentic model. As the most capable 3T-class model in the series, it utilizes large-scale joint reinforcement learning across hundreds of domains. This model is designed to provide significant performance improvements in complex tasks like coding, financial research, and precise 3D creation.
2Technical Architecture Overview
The model features a Mixture-of-Experts architecture with 2.8 trillion total parameters and 104 billion activated parameters. It incorporates 896 experts, selecting 16 per token for efficiency. The system utilizes a 1-million-token context window and replaces traditional attention layers with Gated DeepSeek Sparse Attention to optimize long-context modeling performance during complex operations.
3Multimodal Processing Capabilities
Kimi K3.1 Pro integrates text and image modalities through the MoonViT-V2 vision encoder, which contains 401 million parameters. The model supports quantization-aware training using MXFP4 weights and MXFP8 activations. This multimodal foundation allows the agent to process visual data alongside textual inputs to solve tasks involving document analysis and graphic design.
4Coding and Technical Benchmarks
Performance evaluations involved rigorous testing across benchmarks like Terminal-Bench 4.0, SWE-Marathon, and ProgramBench. The model was assessed using the Claude Code harness, often compared against models like DeepSeek-V4.1-Flash-1017. Results indicate high proficiency in automated software engineering tasks, vulnerability discovery, and complex code generation, with specific scores reported across various industry-standard technical evaluation frameworks.
5Agentic Reasoning Performance
The model demonstrates advanced agentic capabilities in benchmarks such as BioMysteryBench, OfficeQA Pro, and OSWorld V2.1. These tests evaluate the agent's ability to navigate environments, retrieve information, and execute multi-step tool calls. By utilizing specialized harnesses, the model maintains factual accuracy and reasoning depth even when handling large PDF corpora or complex CLI tasks.
6Deployment and API Integration
Developers can access the model via the platform.kimi.ai API, which supports OpenAI and Anthropic-compatible endpoints. The implementation requires preserving the assistant's reasoning content and tool calls in multi-turn conversations. Users can configure the model's thinking effort using the reasoning_effort field, with options for low, high, and max settings to balance speed.
7Evaluation Methodology Standards
All reported results utilize specific configurations, including a temperature of 1.0 and top-p of 0.95, unless otherwise noted. Internal benchmarks like Kimi Code Bench 2.0 and 3.0 Long-horizon provide proprietary insights into performance. Automated graders for these evaluations were calibrated against thousands of human judgments to ensure the reported metrics reflect actual model utility.
8Licensing and Future Outlook
Kimi K3.1 Pro is released under the specific Kimi K3.1 Pro License, with model files available in Safetensors format. This release marks a shift toward more capable, open-weight agentic systems. Organizations should monitor how these 2.8T parameter models influence automated workflows in finance, legal research, and software development sectors moving forward.