kimi-k25

name: kimi-k2.5 description: | Kimi K2.5 Visual Agentic Model routing and configuration. Use when: a task requires visual reasoning, video input processing, multimodal agent orchestration, or deploying Moonshot AI's 1T-parameter agentic model for swarm tasks. role: Specialist Executor intent: Deploy Moonshot AI's 1T-param multimodal agentic model for visual reasoning, video input, and agent swarm tasks. kpis: - visual_reasoning_accuracy - agent_task_completion_rate - cost_per_inference status: active priority: high triggers: - kimi - kimi k2.5 - moonshot - multimodal agent - vision agent - agent swarm - visual reasoning - video input - 1 trillion execution: sequential — classify visual/agentic task → configure K2.5 endpoint → submit input → parse structured output → log metrics dependencies: - HF_TOKEN

Kimi K2.5 — Visual Agentic Model

Moonshot AI's Kimi K2.5 is a 1 trillion parameter multimodal agentic model designed for visual reasoning, video understanding, and autonomous agent swarm coordination. Licensed under Modified MIT.

Model Overview

Capabilities

Visual Reasoning

• Scene understanding and spatial reasoning from images
• Chart/graph interpretation and data extraction
• UI screenshot analysis and element identification
• Document layout parsing (invoices, forms, blueprints)

Video Input

• Frame-by-frame analysis of video clips
• Temporal event detection and sequencing
• Action recognition and scene transition mapping

Agent Swarm Coordination

• Multi-agent task decomposition from visual inputs
• Structured output generation for downstream agent consumption
• Tool-use planning based on visual context

Routing Rules

1. Use Kimi K2.5 for visual reasoning tasks that require understanding spatial relationships, charts, UI layouts, or document structures.
2. Use for video input processing — K2.5 natively handles video frames without requiring external frame extraction.
3. Use for agent swarm orchestration when the task involves visual context that must be decomposed into sub-agent assignments.
4. Do NOT use for text-only tasks — K2.5's 1T parameter cost is not justified for pure text inference. Use GLM-5 or Gemini 3.1 Pro instead.
5. Requires HF_TOKEN — access is via Hugging Face inference endpoints.

When NOT to Use Kimi K2.5

• Pure text generation or reasoning (use GLM-5 or Gemini 3.1 Pro).
• Simple image classification that Vision API or Gemini Flash can handle.
• Tasks where latency is critical — 1T MoE models have higher inference latency.
• Budget-constrained tasks where a smaller multimodal model suffices.

Anti-Patterns

• Routing text-only tasks to K2.5 (massive cost overhead for no visual benefit).
• Using K2.5 for simple image labeling that Gemini Flash handles at a fraction of the cost.
• Not parsing the structured agentic output — K2.5 returns tool-use plans that must be consumed.
• Ignoring the Modified MIT license terms when redistributing model outputs.

Integration Example

import { HfInference } from "@huggingface/inference";

const hf = new HfInference(process.env.HF_TOKEN);

const result = await hf.chatCompletion({
  model: "moonshot-ai/kimi-k2.5",
  messages: [
    {
      role: "user",
      content: [
        { type: "text", text: "Analyze this UI screenshot and list all interactive elements." },
        { type: "image_url", image_url: { url: screenshotUrl } }
      ]
    }
  ]
});