Moonshot AI Releases Kimi K2.5: Open-Source Multimodal Model Advances Visual Agentic Intelligence
By Administrator
Moonshot AI has launched Kimi K2.5, an open-source native multimodal model excelling in agentic benchmarks, vision understanding, coding, and parallel agent execution through its innovative Agent Swarm feature.
Moonshot AI, the Chinese artificial intelligence company, has unveiled Kimi K2.5, described as its most powerful open-source model to date. The release, announced via the official Kimi.ai X account, positions the model as a leader in visual agentic intelligence, combining advanced multimodal capabilities with self-directed parallel agent execution.
Kimi K2.5 builds on the foundation of Kimi K2 through continued pretraining on approximately 15 trillion mixed visual and text tokens. The model adopts a native multimodal architecture, enabling seamless processing of text, images, and videos from a single prompt. It is equipped with a Mixture-of-Experts (MoE) design featuring 1 trillion total parameters and 32 billion activated parameters, supported by a 256K token context length.
Benchmark Performance Highlights
Kimi K2.5 demonstrates state-of-the-art (SOTA) results across several categories, particularly in agentic, vision, coding, and video understanding tasks.
In agentic benchmarks, the model achieves 50.2% on HLE full set with tools (global SOTA) and 74.9% on BrowseComp (with context management). When utilizing Agent Swarm, performance increases to 78.4% on BrowseComp and 79.0% on WideSearch (item-f1).
For vision and video benchmarks, Kimi K2.5 scores 78.5% on MMMU Pro (open-source SOTA), 86.6% on VideoMMMU, and 84.2% on MathVision. These results position it competitively against leading proprietary models, though some competitors such as GPT-5.2 and Gemini 3 Pro report slightly higher scores in select categories.
In coding benchmarks, the model attains 76.8% on SWE-bench Verified (open-source SOTA), 85.0% on LiveCodeBench (v6), and 50.8% on Terminal-Bench 2.0. These figures underscore its strength in software engineering tasks and visual coding applications.
Key Features and Innovations
A standout capability is "Code with Taste", which allows the model to transform chats, images, or videos into aesthetic websites featuring expressive motion and interactive elements. This includes reconstructing websites from video demonstrations, generating front-end interfaces with animations, and performing visual reasoning tasks such as solving mazes via algorithmic approaches like BFS.
The Agent Swarm (Beta) feature represents a significant advancement in agentic execution. It enables the model to self-direct up to 100 sub-agents and coordinate up to 1,500 tool calls in parallel workflows. This approach, trained via Parallel-Agent Reinforcement Learning (PARL), reduces execution time by up to 4.5× compared to single-agent setups by decomposing complex tasks and minimizing critical path latency.
Kimi K2.5 also supports enhanced office productivity tasks, long-form content generation (up to 10,000-word outputs), and multimodal reasoning for applications including document annotation, financial modeling, and LaTeX processing.
Availability and Access
Kimi K2.5 is immediately accessible in chat and agent modes on kimi.com and the Kimi app, with Agent Swarm available in beta for high-tier users. The model supports thinking mode for step-by-step reasoning and instant mode for faster responses.
Developers can access Kimi K2.5 via the Moonshot AI API at platform.moonshot.ai, which offers OpenAI-compatible endpoints for multimodal inputs. Model weights and code are openly available on Hugging Face under a Modified MIT License, facilitating community deployment and fine-tuning.
For production-grade coding workflows, Moonshot recommends pairing Kimi K2.5 with Kimi Code, an open-source terminal tool integrated with IDEs like VS Code.
Context and Implications
The release of Kimi K2.5 strengthens Moonshot AI's position in the competitive global AI landscape, particularly among open-source contributors. By providing free access to API endpoints and model weights, the company enables broader experimentation and adoption of its multimodal and agentic technologies.
Benchmarks were conducted under standardized conditions (temperature=1.0, top-p=0.95), with vision tasks using up to 64K tokens. Some comparative scores from competitors were re-evaluated for consistency. The model includes experimental features such as video chat support, limited to the official API.
Moonshot AI continues to emphasize open-source development as a means to advance artificial intelligence capabilities for diverse applications.