MoonshotAI: Kimi K2.5
Model Type
Open Weight Model
1T parameters
Recommended Use Cases
Text Generation
Try Kimi K2.5
Overview
Kimi K2.5 is Moonshot AI's most powerful open-source model, released in January 2026, featuring native multimodal capabilities (text, image, video), Agent Swarm orchestration, and state-of-the-art coding with vision. K2.5 builds on Kimi K2 with continued pretraining over approximately 15 trillion mixed visual and text tokens. It delivers native multimodal understanding, generates code from visual specifications, and introduces a self-directed Agent Swarm paradigm that can coordinate up to 100 sub-agents executing parallel workflows across up to 1,500 tool calls.
Technical Specifications
| Specification | Value |
|---|---|
| Total Parameters | 1T |
| Active Parameters | 32B |
| Architecture | MoE with MLA (Multi-head Latent Attention) |
| Vision Encoder | MoonViT (400M parameters) |
| Context Length | 256K tokens |
| Training Data | ~15T mixed visual and text tokens |
Modes
Kimi K2.5 supports four distinct operating modes:
- K2.5 Instant: Direct responses without reasoning traces
- K2.5 Thinking: Includes reasoning traces with extended chain-of-thought
- K2.5 Agent: Single-agent task execution with tool use
- K2.5 Agent Swarm (Beta): Multi-agent coordination for complex tasks
Capabilities
- Frontend development from visual designs (HTML, React, responsive web applications)
- Video-to-code generation from workflow recordings
- Office productivity (Word annotations, Excel pivot tables, LaTeX in PDFs)
- Long-form outputs (10,000-word papers, 100-page documents)
- Agentic search with web browsing and code interpreter
Role in Series
Moonshot AI's Kimi models offer different capabilities:
- Kimi Linear: Most efficient, 75% less KV cache, 6x throughput
- Kimi K2 0905: Agentic coding specialist, 256K context
- Kimi K2.5: Maximum capability with vision and Agent Swarm (this model)