Paper 2602.02276
Kimi K2.5: Visual Agentic Intelligence
- Published
- Feb 2026
- Research lab
- Moonshot AI
- Citations
- 313
- GitHub
- 2.3K stars
01 In brief
Summary
Kimi K2.5 is an open-source multimodal agentic model that jointly optimizes text and vision through techniques including joint pre-training, zero-vision SFT, and joint reinforcement learning.
It introduces Agent Swarm, a parallel agent orchestration framework using Parallel-Agent Reinforcement Learning (PARL) to decompose tasks into heterogeneous sub-problems executed concurrently.
Evaluations show state-of-the-art results across coding, vision, reasoning, and agentic tasks, with Agent Swarm reducing latency by up to 4.5× and improving WideSearch item-level F1 from 72.8% to 79.0% over single-agent baselines.
The model achieves 96.1% on AIME 2025, 76.8% on SWE-Bench Verified, 78.4% on BrowseComp with Agent Swarm, and 63.3% on OSWorld-Verified.
Post-trained checkpoints are released on Hugging Face.
Key findings include that early vision fusion with lower ratios yields better results, text-only SFT activates visual capabilities, and visual RL improves textual performance (e.g., MMLU-Pro from 84.7% to 86.4%).
The architecture uses MoonViT-3D for native-resolution image and video processing, and training employs the Decoupled Encoder Process for efficiency.
The Toggle heuristic reduces output tokens by 25–30% with negligible performance impact.
The model is built on Kimi K2 with 1.04 trillion total parameters and 32 billion activated parameters.
02 From the paper
Abstract
We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.