Cloudflare Workers Ai

---
name: cloudflare-workers-ai
description: >-
  Reference for Cloudflare Workers AI, running AI models on serverless GPUs through the
  env.AI.run() binding. Covers the AI binding in wrangler.jsonc, model selection (@cf/meta/llama,
  Flux and Stable Diffusion image generation, BGE embeddings, vision models), streaming responses,
  AI Gateway caching, RAG with Vectorize, Vercel AI SDK integration and OpenAI-compatible
  endpoints, plus rate limits, neuron pricing and a production checklist. Use when implementing AI
  inference in a Cloudflare Worker, generating text or images at the edge, configuring Workers AI
  bindings, adding AI streaming, wiring up AI Gateway, building embeddings or RAG on Workers, or
  debugging AI_ERROR, rate limit errors, model not found, model unavailable, token limit exceeded,
  neurons exceeded, quota exceeded or streaming failures.
license: MIT
---

# Cloudflare Workers AI - Complete Reference

Production-ready knowledge domain for building AI-powered applications with Cloudflare Workers AI.

**Status**: Production Ready ✅
**Last Updated**: 2025-10-21
**Dependencies**: cloudflare-worker-base (for Worker setup)
**Latest Versions**: wrangler@4.43.0, @cloudflare/workers-types@4.20251014.0

---

## Table of Contents

1. [Quick Start (5 minutes)](#quick-start-5-minutes)
2. [Workers AI API Reference](#workers-ai-api-reference)
3. [Model Selection Guide](#model-selection-guide)
4. [Common Patterns](#common-patterns)
5. [AI Gateway Integration](#ai-gateway-integration)
6. [Rate Limits & Pricing](#rate-limits--pricing)
7. [Production Checklist](#production-checklist)

---

## Quick Start (5 minutes)

### 1. Add AI Binding

**wrangler.jsonc:**
```jsonc
{
  "ai": {
    "binding": "AI"
  }
}
```

### 2. Run Your First Model

```typescript
export interface Env {
  AI: Ai;
}

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    const response = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
      prompt: 'What is Cloudflare?',
    });

    return Response.json(response);
  },
};
```

### 3. Add Streaming (Recommended)

```typescript
const stream = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
  messages: [{ role: 'user', content: 'Tell me a story' }],
  stream: true, // Always use streaming for text generation!
});

return new Response(stream, {
  headers: { 'content-type': 'text/event-stream' },
});
```

**Why streaming?**
- Prevents buffering large responses in memory
- Faster time-to-first-token
- Better user experience for long-form content
- Avoids Worker timeout issues

---

## Workers AI API Reference

### `env.AI.run()`

Run an AI model inference.

**Signature:**
```typescript
async env.AI.run(
  model: string,
  inputs: ModelInputs,
  options?: { gateway?: { id: string; skipCache?: boolean } }
): Promise<ModelOutput | ReadableStream>
```

**Parameters:**

- `model` (string, required) - Model ID (e.g., `@cf/meta/llama-3.1-8b-instruct`)
- `inputs` (object, required) - Model-specific inputs
- `options` (object, optional) - Additional options
  - `gateway` (object) - AI Gateway configuration
    - `id` (string) - Gateway ID
    - `skipCache` (boolean) - Skip AI Gateway cache

**Returns:**

- Non-streaming: `Promise<ModelOutput>` - JSON response
- Streaming: `ReadableStream` - Server-sent events stream

---

### Text Generation Models

**Input Format:**
```typescript
{
  messages?: Array<{ role: 'system' | 'user' | 'assistant'; content: string }>;
  prompt?: string; // Deprecated, use messages
  stream?: boolean; // Default: false
  max_tokens?: number; // Max tokens to generate
  temperature?: number; // 0.0-1.0, default varies by model
  top_p?: number; // 0.0-1.0
  top_k?: number;
}
```

**Output Format (Non-Streaming):**
```typescript
{
  response: string; // Generated text
}
```

**Example:**
```typescript
const response = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'What is TypeScript?' },
  ],
  stream: false,
});

console.log(response.response);
```

---

### Text Embeddings Models

**Input Format:**
```typescript
{
  text: string | string[]; // Single text or array of texts
}
```

**Output Format:**
```typescript
{
  shape: number[]; // [batch_size, embedding_dimensions]
  data: number[][]; // Array of embedding vectors
}
```

**Example:**
```typescript
const embeddings = await env.AI.run('@cf/baai/bge-base-en-v1.5', {
  text: ['Hello world', 'Cloudflare Workers'],
});

console.log(embeddings.shape); // [2, 768]
console.log(embeddings.data[0]); // [0.123, -0.456, ...]
```

---

### Image Generation Models

**Input Format:**
```typescript
{
  prompt: string; // Text description
  num_steps?: number; // Default: 20
  guidance?: number; // CFG scale, default: 7.5
  strength?: number; // For img2img, default: 1.0
  image?: number[][]; // For img2img (base64 or array)
}
```

**Output Format:**
- Binary image data (PNG/JPEG)

**Example:**
```typescript
const imageStream = await env.AI.run('@cf/black-forest-labs/flux-1-schnell', {
  prompt: 'A beautiful sunset over mountains',
});

return new Response(imageStream, {
  headers: { 'content-type': 'image/png' },
});
```

---

### Vision Models

**Input Format:**
```typescript
{
  messages: Array<{
    role: 'user' | 'assistant';
    content: Array<{ type: 'text' | 'image_url'; text?: string; image_url?: { url: string } }>;
  }>;
}
```

**Example:**
```typescript
const response = await env.AI.run('@cf/meta/llama-3.2-11b-vision-instruct', {
  messages: [
    {
      role: 'user',
      content: [
        { type: 'text', text: 'What is in this image?' },
        { type: 'image_url', image_url: { url: 'data:image/png;base64,iVBOR...' } },
      ],
    },
  ],
});
```

---

## Model Selection Guide

### Text Generation (LLMs)

| Model | Best For | Rate Limit | Size |
|-------|----------|------------|------|
| `@cf/meta/llama-3.1-8b-instruct` | General purpose, fast | 300/min | 8B |
| `@cf/meta/llama-3.2-1b-instruct` | Ultra-fast, simple tasks | 300/min | 1B |
| `@cf/qwen/qwen1.5-14b-chat-awq` | High quality, complex reasoning | 150/min | 14B |
| `@cf/deepseek-ai/deepseek-r1-distill-qwen-32b` | Coding, technical content | 300/min | 32B |
| `@hf/thebloke/mistral-7b-instruct-v0.1-awq` | Fast, efficient | 400/min | 7B |

### Text Embeddings

| Model | Dimensions | Best For | Rate Limit |
|-------|-----------|----------|------------|
| `@cf/baai/bge-base-en-v1.5` | 768 | General purpose RAG | 3000/min |
| `@cf/baai/bge-large-en-v1.5` | 1024 | High accuracy search | 1500/min |
| `@cf/baai/bge-small-en-v1.5` | 384 | Fast, low storage | 3000/min |

### Image Generation

| Model | Best For | Rate Limit | Speed |
|-------|----------|------------|-------|
| `@cf/black-forest-labs/flux-1-schnell` | High quality, photorealistic | 720/min | Fast |
| `@cf/stabilityai/stable-diffusion-xl-base-1.0` | General purpose | 720/min | Medium |
| `@cf/lykon/dreamshaper-8-lcm` | Artistic, stylized | 720/min | Fast |

### Vision Models

| Model | Best For | Rate Limit |
|-------|----------|------------|
| `@cf/meta/llama-3.2-11b-vision-instruct` | Image understanding | 720/min |
| `@cf/unum/uform-gen2-qwen-500m` | Fast image captioning | 720/min |

---

## Common Patterns

### Pattern 1: Chat Completion with History

```typescript
app.post('/chat', async (c) => {
  const { messages } = await c.req.json<{
    messages: Array<{ role: string; content: string }>;
  }>();

  const response = await c.env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
    messages,
    stream: true,
  });

  return new Response(response, {
    headers: { 'content-type': 'text/event-stream' },
  });
});
```

---

### Pattern 2: RAG (Retrieval Augmented Generation)

```typescript
// Step 1: Generate embeddings
const embeddings = await env.AI.run('@cf/baai/bge-base-en-v1.5', {
  text: [userQuery],
});

const vector = embeddings.data[0];

// Step 2: Search Vectorize
const matches = await env.VECTORIZE.query(vector, { topK: 3 });

// Step 3: Build context from matches
const context = matches.matches.map((m) => m.metadata.text).join('\n\n');

// Step 4: Generate response with context
const response = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
  messages: [
    {
      role: 'system',
      content: `Answer using this context:\n${context}`,
    },
    { role: 'user', content: userQuery },
  ],
  stream: true,
});

return new Response(response, {
  headers: { 'content-type': 'text/event-stream' },
});
```

---

### Pattern 3: Structured Output with Zod

```typescript
import { z } from 'zod';

const RecipeSchema = z.object({
  name: z.string(),
  ingredients: z.array(z.string()),
  instructions: z.array(z.string()),
  prepTime: z.number(),
});

app.post('/recipe', async (c) => {
  const { dish } = await c.req.json<{ dish: string }>();

  const response = await c.env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
    messages: [
      {
        role: 'user',
        content: `Generate a recipe for ${dish}. Return ONLY valid JSON matching this schema: ${JSON.stringify(RecipeSchema.shape)}`,
      },
    ],
  });

  // Parse and validate
  const recipe = RecipeSchema.parse(JSON.parse(response.response));

  return c.json(recipe);
});
```

---

### Pattern 4: Image Generation + R2 Storage

```typescript
app.post('/generate-image', async (c) => {
  const { prompt } = await c.req.json<{ prompt: string }>();

  // Generate image
  const imageStream = await c.env.AI.run('@cf/black-forest-labs/flux-1-schnell', {
    prompt,
  });

  const imageBytes = await new Response(imageStream).bytes();

  // Store in R2
  const key = `images/${Date.now()}.png`;
  await c.env.BUCKET.put(key, imageBytes, {
    httpMetadata: { contentType: 'image/png' },
  });

  return c.json({
    success: true,
    url: `https://your-domain.com/${key}`,
  });
});
```

---

## AI Gateway Integration

AI Gateway provides caching, logging, and analytics for AI requests.

**Setup:**
```typescript
const response = await env.AI.run(
  '@cf/meta/llama-3.1-8b-instruct',
  { prompt: 'Hello' },
  {
    gateway: {
      id: 'my-gateway', // Your gateway ID
      skipCache: false, // Use cache
    },
  }
);
```

**Benefits:**
- ✅ **Cost Tracking** - Monitor neurons usage per request
- ✅ **Caching** - Reduce duplicate inference costs
- ✅ **Logging** - Debug and analyze AI requests
- ✅ **Rate Limiting** - Additional layer of protection
- ✅ **Analytics** - Request patterns and performance

**Access Gateway Logs:**
```typescript
const gateway = env.AI.gateway('my-gateway');
const logId = env.AI.aiGatewayLogId;

// Send feedback
await gateway.patchLog(logId, {
  feedback: { rating: 1, comment: 'Great response' },
});
```

---

## Rate Limits & Pricing

### Rate Limits (per minute)

| Task Type | Default Limit | Notes |
|-----------|---------------|-------|
| **Text Generation** | 300/min | Some fast models: 400-1500/min |
| **Text Embeddings** | 3000/min | BGE-large: 1500/min |
| **Image Generation** | 720/min | All image models |
| **Vision Models** | 720/min | Image understanding |
| **Translation** | 720/min | M2M100, Opus MT |
| **Classification** | 2000/min | Text classification |
| **Speech Recognition** | 720/min | Whisper models |

### Pricing (Neurons-Based)

**Free Tier:**
- 10,000 neurons per day
- Resets daily at 00:00 UTC

**Paid Tier:**
- $0.011 per 1,000 neurons
- 10,000 neurons/day included
- Unlimited usage above free allocation

**Example Costs:**

| Model | Input (1M tokens) | Output (1M tokens) |
|-------|-------------------|-------------------|
| Llama 3.2 1B | $0.027 | $0.201 |
| Llama 3.1 8B | $0.088 | $0.606 |
| BGE-base embeddings | $0.005 | N/A |
| Flux image generation | ~$0.011/image | N/A |

---

## Production Checklist

### Before Deploying

- [ ] **Enable AI Gateway** for cost tracking and logging
- [ ] **Implement streaming** for all text generation endpoints
- [ ] **Add rate limit retry** with exponential backoff
- [ ] **Validate input length** to prevent token limit errors
- [ ] **Set appropriate timeouts** (Workers: 30s CPU default, 5m max)
- [ ] **Monitor neurons usage** in Cloudflare dashboard
- [ ] **Test error handling** for model unavailable, rate limits
- [ ] **Add input sanitization** to prevent prompt injection
- [ ] **Configure CORS** if using from browser
- [ ] **Plan for scale** - upgrade to Paid plan if needed

### Error Handling

```typescript
async function runAIWithRetry(
  env: Env,
  model: string,
  inputs: any,
  maxRetries = 3
): Promise<any> {
  let lastError: Error;

  for (let i = 0; i < maxRetries; i++) {
    try {
      return await env.AI.run(model, inputs);
    } catch (error) {
      lastError = error as Error;
      const message = lastError.message.toLowerCase();

      // Rate limit - retry with backoff
      if (message.includes('429') || message.includes('rate limit')) {
        const delay = Math.pow(2, i) * 1000; // Exponential backoff
        await new Promise((resolve) => setTimeout(resolve, delay));
        continue;
      }

      // Other errors - throw immediately
      throw error;
    }
  }

  throw lastError!;
}
```

### Monitoring

```typescript
app.use('*', async (c, next) => {
  const start = Date.now();

  await next();

  // Log AI usage
  console.log({
    path: c.req.path,
    duration: Date.now() - start,
    logId: c.env.AI.aiGatewayLogId,
  });
});
```

---

## OpenAI Compatibility

Workers AI supports OpenAI-compatible endpoints.

**Using OpenAI SDK:**
```typescript
import OpenAI from 'openai';

const openai = new OpenAI({
  apiKey: env.CLOUDFLARE_API_KEY,
  baseURL: `https://api.cloudflare.com/client/v4/accounts/${env.CLOUDFLARE_ACCOUNT_ID}/ai/v1`,
});

// Chat completions
const completion = await openai.chat.completions.create({
  model: '@cf/meta/llama-3.1-8b-instruct',
  messages: [{ role: 'user', content: 'Hello!' }],
});

// Embeddings
const embeddings = await openai.embeddings.create({
  model: '@cf/baai/bge-base-en-v1.5',
  input: 'Hello world',
});
```

**Endpoints:**
- `/v1/chat/completions` - Text generation
- `/v1/embeddings` - Text embeddings

---

## Vercel AI SDK Integration

```bash
npm install workers-ai-provider ai
```

```typescript
import { createWorkersAI } from 'workers-ai-provider';
import { generateText, streamText } from 'ai';

const workersai = createWorkersAI({ binding: env.AI });

// Generate text
const result = await generateText({
  model: workersai('@cf/meta/llama-3.1-8b-instruct'),
  prompt: 'Write a poem',
});

// Stream text
const stream = streamText({
  model: workersai('@cf/meta/llama-3.1-8b-instruct'),
  prompt: 'Tell me a story',
});
```

---

## Limits Summary

| Feature | Limit |
|---------|-------|
| Concurrent requests | No hard limit (rate limits apply) |
| Max input tokens | Varies by model (typically 2K-128K) |
| Max output tokens | Varies by model (typically 512-2048) |
| Streaming chunk size | ~1 KB |
| Image size (output) | ~5 MB |
| Request timeout | Workers timeout applies (30s default, 5m max CPU) |
| Daily free neurons | 10,000 |
| Rate limits | See "Rate Limits & Pricing" section |

---

## References

- [Workers AI Docs](https://developers.cloudflare.com/workers-ai/)
- [Models Catalog](https://developers.cloudflare.com/workers-ai/models/)
- [AI Gateway](https://developers.cloudflare.com/ai-gateway/)
- [Pricing](https://developers.cloudflare.com/workers-ai/platform/pricing/)
- [Limits](https://developers.cloudflare.com/workers-ai/platform/limits/)
- [REST API](https://developers.cloudflare.com/workers-ai/get-started/rest-api/)