Kimi K2.5 on Workers AI: Frontier MoE for $0.60/M

A frontier-scale MoE model with 256k context, multi-turn tool calling, and vision — now one API call away on Cloudflare's global network, at $0.60 per million input tokens.

What Is Kimi K2.5?

Kimi K2.5 is a frontier-scale large language model built by Moonshot AI. It uses a Mixture of Experts (MoE) architecture — the same design pattern behind models like Mixtral and DeepSeek — where only a subset of the model's parameters activate for each token, delivering high capability at lower inference cost.

The model is fully open-source and now available on Cloudflare Workers AI under the model ID @cf/moonshotai/kimi-k2.5. It ships with a 256,000-token context window — large enough to ingest entire codebases, lengthy documents, or multi-turn agent conversations without truncation.

Multi-Turn Tool Calling and Structured Output

Kimi K2.5 supports multi-turn tool calling with parallel tool use. You define tools in the standard OpenAI-compatible format, and the model can call multiple tools in a single response. After receiving the results, it reasons over them and decides on the next action — or returns a final answer.

It also supports vision inputs (image understanding) and structured outputs via JSON mode and JSON Schema. You can force the model to return a specific JSON shape — useful for building reliable pipelines where the output feeds directly into another system.

77% Cost Savings: Cloudflare's Own Numbers

Pricing

The cached input price is where the economics get interesting. If your agent uses a consistent system prompt or feeds the same document context across multiple calls, prefix caching drops your input cost to $0.10 per million tokens — an 83% discount on the already-low base price.

Prefix Caching and Session Affinity

Prefix caching works automatically when consecutive requests share a common prefix (system prompt, prior conversation turns). Workers AI caches the KV state for the shared prefix and reuses it, cutting both cost and latency.

To maximize cache hit rates in multi-turn conversations, set the x-session-affinity header. This routes requests from the same session to the same inference node, so the cached prefix is available locally instead of requiring a cache lookup across the network.

// REST API call with session affinity for prefix caching
const response = await fetch(
  'https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/moonshotai/kimi-k2.5',
  {
    method: 'POST',
    headers: {
      'Authorization': 'Bearer {api_token}',
      'Content-Type': 'application/json',
      'x-session-affinity': 'session-abc-123'  // pin to same node
    },
    body: JSON.stringify({
      messages: [
        { role: 'system', content: 'You are a database optimization assistant.' },
        { role: 'user', content: 'Analyze this slow query...' }
      ]
    })
  }
);

Code Examples

Workers Binding

The simplest path: bind Workers AI in your wrangler.toml and call the model directly from your Worker. No API keys, no external HTTP calls — the binding handles auth and routing automatically.

// wrangler.toml
// [ai]
// binding = "AI"

export default {
  async fetch(request, env) {
    const response = await env.AI.run(
      '@cf/moonshotai/kimi-k2.5',
      {
        messages: [
          {
            role: 'system',
            content: 'You are a helpful assistant specialized in SQL optimization for SQLite and D1.'
          },
          {
            role: 'user',
            content: 'Rewrite this query to use a covering index: SELECT name, email FROM users WHERE created_at > date("now", "-7 days") ORDER BY created_at DESC'
          }
        ],
        max_tokens: 2048
      }
    );

    return Response.json(response);
  }
};

Async Batch Processing

For non-latency-sensitive workloads — bulk classification, document summarization, data extraction — the Async API lets you submit requests and poll for results. This is pull-based batch processing: you submit the job, get a task ID, and fetch the result when it is ready.

// Submit an async task
const task = await fetch(
  'https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/moonshotai/kimi-k2.5',
  {
    method: 'POST',
    headers: {
      'Authorization': 'Bearer {api_token}',
      'Content-Type': 'application/json',
      'cf-aig-async': 'true'  // enable async mode
    },
    body: JSON.stringify({
      messages: [
        { role: 'user', content: 'Summarize the following 50-page document...' }
      ]
    })
  }
);

const { taskId } = await task.json();

// Poll for the result later
const result = await fetch(
  'https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/tasks/{taskId}',
  { headers: { 'Authorization': 'Bearer {api_token}' } }
);

Now the Default in Agents SDK

Kimi K2.5 is now the default model in Cloudflare's Agents SDK starter template. When you scaffold a new agent with the SDK, it comes pre-configured with Kimi K2.5 — ready for tool calling, multi-turn reasoning, and structured output out of the box.

This signals where Cloudflare sees the model fitting: not just as another inference endpoint, but as the recommended brain for agent workloads on their platform.

The Full Agent Stack on One Platform

When to Use Kimi K2.5 on Workers AI

The Bottom Line

Kimi K2.5 on Workers AI is not just another model added to a catalog. It is a platform-level move — Cloudflare eating into the inference market the same way they ate into CDN, DNS, and serverless. The 256k context window, parallel tool calling, and prefix caching make it purpose-built for the agentic future. The 77% cost savings from their own internal use case prove it works at scale.

If you are building on the Cloudflare Workers stack, Kimi K2.5 is the obvious default for agent workloads. If you are evaluating inference providers, the combination of pricing, context length, and zero-egress platform integration is hard to beat.

Already running D1 databases? Download MyD1 to browse and query them visually — and let the AI Agent handle the complex queries for you.

Sources: Cloudflare blog announcement · Kimi K2.5 model documentation

Related: AWS EC2 vs Cloudflare Workers Stack · Which AI Model Writes the Best SQL for D1? · Build a Full-Stack App on Cloudflare for Free