News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgvidfnKimi K2.5

Kimi K2.5

Open-source multimodal agentic model with vision and reasoning.

Moonshot AI released Kimi K2.5, an open-source mixture-of-experts model that processes text, images and video. It uses 32 billion active parameters and supports agent swarms for complex tasks, offering both instant and deep thinking modes via API.

Share on WhatsApp

  • Trained on 15 trillion mixed visual and text tokens.
  • Uses 32 billion activated parameters in MoE architecture.
  • Supports image and video input natively.
  • Features agent swarm capability for parallel task execution.
  • Offers both instant and deep thinking modes.

What it is

Kimi K2.5 is a native multimodal agentic model developed by Moonshot AI. It builds upon the Kimi-K2-Base architecture through continual pretraining on roughly 15 trillion tokens that mix visual and textual data. The model integrates vision and language understanding with advanced capabilities for using tools and acting as an agent. It features both instant response modes and deeper thinking modes, allowing it to handle conversational tasks as well as complex agentic workflows. The system is designed to process images and video inputs alongside text, enabling it to reason across different types of media.

How it came about

The model was built on top of the Kimi-K2-Base foundation. Moonshot AI conducted continual pretraining to expand its capabilities, specifically focusing on mixing visual and text tokens. This development aimed to create a system that could seamlessly integrate visual knowledge with language understanding. The release includes updates to the chat template and removes default system prompts that previously caused confusion. The model represents a shift from single-agent scaling to more coordinated execution schemes.

Inside the model

Kimi K2.5 uses a Mixture-of-Experts architecture with 61 layers, including one dense layer. It has 384 experts in total, selecting eight per token, with one shared expert. The attention mechanism is MLA, and it uses SwiGLU activation. A MoonViT vision encoder with 400 million parameters handles visual inputs. The model supports a context length of 256,000 tokens. It employs native INT4 quantization for efficiency. The design allows for dynamic instantiation of domain-specific agents to decompose complex tasks into parallel sub-tasks.

What it can do

The model excels in visual knowledge and cross-modal reasoning. It can generate code from visual specifications like user interface designs or video workflows. In benchmarks, it performs tasks requiring deep tool use, such as web browsing and code interpretation. Evaluations show scores on various tests, including HLE and WorldVQA, often comparing favourably against other leading models when tools are enabled. It supports an agent swarm mode for handling wide search and complex multi-step reasoning tasks.

How to use it

Moonshot AI provides access to Kimi K2.5 via its official API, which is compatible with OpenAI and Anthropic standards. The weights are open-source, allowing for local deployment using engines like vLLM or SGLang. Users can verify deployments with a provided vendor verifier. The model requires transformers version 4.57.1 or higher. Video chat is currently an experimental feature available only through the official API. Recommended settings include a temperature of 1.0 for thinking mode and 0.6 for instant mode.

Questions people ask about Kimi K2.5

What types of media can Kimi K2.5 process?

Kimi K2.5 is a native multimodal model that accepts text, images, and video as inputs. It uses a dedicated MoonViT vision encoder to process visual data, allowing it to perform cross-modal reasoning and generate code from visual specifications like UI designs.

How does the agent swarm feature work?

The agent swarm feature allows the model to decompose complex tasks into parallel sub-tasks. It dynamically instantiates domain-specific agents to execute these sub-tasks in a coordinated manner, moving beyond single-agent scaling to handle more intricate workflows efficiently.

Is Kimi K2.5 available for local use?

Yes, the model weights are open-source. It can be deployed locally using inference engines such as vLLM or SGLang. The minimum required version for the transformers library is 4.57.1, and it supports native INT4 quantization for efficient running.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Kimi K2.5 yet, so the facts on this page come from the sources listed above: huggingface.co/moonshotai/Kimi-K2.5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.