Cloudflare Workers Ai

Cloudflare Workers AI

Status: Production Ready ✅ Last Updated: 2026-01-21 Dependencies: cloudflare-worker-base (for Worker setup) Latest Versions: wrangler@4.58.0, @cloudflare/workers-types@4.20260109.0, workers-ai-provider@3.0.2

Recent Updates (2025):

• April 2025 - Performance: Llama 3.3 70B 2-4x faster (speculative decoding, prefix caching), BGE embeddings 2x faster
• April 2025 - Breaking Changes: max_tokens now correctly defaults to 256 (was not respected), BGE pooling parameter (cls NOT backwards compatible with mean)
• 2025 - New Models (14): Mistral 3.1 24B (vision+tools), Gemma 3 12B (128K context), EmbeddingGemma 300M, Llama 4 Scout, GPT-OSS 120B/20B, Qwen models (QwQ 32B, Coder 32B), Leonardo image gen, Deepgram Aura 2, Whisper v3 Turbo, IBM Granite, Nova 3
• 2025 - Platform: Context windows API change (tokens not chars), unit-based pricing with per-model granularity, workers-ai-provider v3.0.2 (AI SDK v5), LoRA rank up to 32 (was 8), 100 adapters per account
• October 2025: Model deprecations (use Llama 4, GPT-OSS instead)

---

Quick Start (5 Minutes)

// 1. Add AI binding to wrangler.jsonc
{ "ai": { "binding": "AI" } }

// 2. Run model with streaming (recommended)
export default {
  async fetch(request: Request, env: Env): Promise {
    const stream = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
      messages: [{ role: 'user', content: 'Tell me a story' }],
      stream: true, // Always stream for text generation!
    });

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

Why streaming? Prevents buffering in memory, faster time-to-first-token, avoids Worker timeout issues.

---

Known Issues Prevention

This skill prevents 7 documented issues:

Issue #1: Context Window Validation Changed to Tokens (February 2025)

Error: "Exceeded character limit" despite model supporting larger context Source: Cloudflare Changelog Why It Happens: Before February 2025, Workers AI validated prompts using a hard 6144 character limit, even for models with larger token-based context windows (e.g., Mistral with 32K tokens). After the update, validation switched to token-based counting. Prevention: Calculate tokens (not characters) when checking context window limits.

import { encode } from 'gpt-tokenizer'; // or model-specific tokenizer

const tokens = encode(prompt);
const contextWindow = 32768; // Model's max tokens (check docs)
const maxResponseTokens = 2048;

if (tokens.length + maxResponseTokens > contextWindow) {
  throw new Error(`Prompt exceeds context window: ${tokens.length} tokens`);
}

const response = await env.AI.run('@cf/mistral/mistral-7b-instruct-v0.2', {
  messages: [{ role: 'user', content: prompt }],
  max_tokens: maxResponseTokens,
});

Issue #2: Neuron Consumption Discrepancies in Dashboard

Error: Dashboard neuron usage significantly exceeds expected token-based calculations Source: Cloudflare Community Discussion Why It Happens: Users report dashboard showing hundred-million-level neuron consumption for K-level token usage, particularly with AutoRAG features and certain models. The discrepancy between expected neuron consumption (based on pricing docs) and actual dashboard metrics is not fully documented. Prevention: Monitor neuron usage via AI Gateway logs and correlate with requests. File support ticket if consumption significantly exceeds expectations.

// Use AI Gateway for detailed request logging
const response = await env.AI.run(
  '@cf/meta/llama-3.1-8b-instruct',
  { messages: [{ role: 'user', content: query }] },
  { gateway: { id: 'my-gateway' } }
);

// Monitor dashboard at: https://dash.cloudflare.com → AI → Workers AI
// Compare neuron usage with token counts
// File support ticket with details if discrepancy persists

Issue #3: AI Binding Requires Remote or Latest Tooling in Local Dev

Error: "MiniflareCoreError: wrapped binding module can't be resolved (internal modules only)" Source: GitHub Issue #6796 Why It Happens: When using Workers AI bindings with Miniflare in local development (particularly with custom Vite plugins), the AI binding requires external workers that aren't properly exposed by older unstable_getMiniflareWorkerOptions. The error occurs when Miniflare can't resolve the internal AI worker module. Prevention: Use remote bindings for AI in local dev, or update to latest @cloudflare/vite-plugin.

// wrangler.jsonc - Option 1: Use remote AI binding in local dev
{
  "ai": { "binding": "AI" },
  "dev": {
    "remote": true // Use production AI binding locally
  }
}

# Option 2: Update to latest tooling
npm install -D @cloudflare/vite-plugin@latest

# Option 3: Use wrangler dev instead of custom Miniflare
npm run dev

Issue #4: Flux Image Generation NSFW Filter False Positives

Error: "AiError: Input prompt contains NSFW content (code 3030)" for innocent prompts Source: Cloudflare Community Discussion Why It Happens: Flux image generation models (@cf/black-forest-labs/flux-1-schnell) sometimes trigger false positive NSFW content errors even with innocent single-word prompts like "hamburger". The NSFW filter can be overly sensitive without context. Prevention: Add descriptive context around potential trigger words instead of using single-word prompts.

// ❌ May trigger error 3030
const response = await env.AI.run('@cf/black-forest-labs/flux-1-schnell', {
  prompt: 'hamburger', // Single word triggers filter
});

// ✅ Add context to avoid false positives
const response = await env.AI.run('@cf/black-forest-labs/flux-1-schnell', {
  prompt: 'A photo of a delicious large hamburger on a plate with lettuce and tomato',
  num_steps: 4,
});

Issue #5: Image Generation Error 1000 - Missing num_steps Parameter

Error: "Error: unexpected type 'int32' with value 'undefined' (code 1000)" Source: Cloudflare Community Discussion Why It Happens: Image generation API calls return error code 1000 when the num_steps parameter is not provided, even though documentation suggests it's optional. The parameter is actually required for most Flux models. Prevention: Always include num_steps: 4 for image generation models (typically 4 for Flux Schnell).

// ✅ Always include num_steps for image generation
const image = await env.AI.run('@cf/black-forest-labs/flux-1-schnell', {
  prompt: 'A beautiful sunset over mountains',
  num_steps: 4, // Required - typically 4 for Flux Schnell
});

// Note: FLUX.2 [klein] 4B has fixed steps=4 (cannot be adjusted)

Issue #6: Zod v4 Incompatibility with Structured Output Tools

Error: Syntax errors and failed transpilation when using Stagehand with Zod v4 Source: GitHub Issue #10798 Why It Happens: Stagehand (browser automation) and some structured output examples in Workers AI fail with Zod v4 (now default). The underlying zod-to-json-schema library doesn't yet support Zod v4, causing transpilation failures. Prevention: Pin Zod to v3 until zod-to-json-schema supports v4.

# Install Zod v3 specifically
npm install zod@3

# Or pin in package.json
{
  "dependencies": {
    "zod": "~3.23.8" // Pin to v3 for compatibility
  }
}

Issue #7: AI Gateway Cache Headers for Per-Request Control

Not an error, but important feature: AI Gateway supports per-request cache control via HTTP headers for custom TTL, cache bypass, and custom cache keys beyond dashboard defaults. Source: AI Gateway Caching Documentation Use When: You need different caching behavior for different requests (e.g., 1 hour for expensive queries, skip cache for real-time data). Implementation: See AI Gateway Integration section below for header usage.

---

API Reference

env.AI.run(
  model: string,
  inputs: ModelInputs,
  options?: { gateway?: { id: string; skipCache?: boolean } }
): Promise

---

Model Selection Guide (Updated 2025)

Text Generation (LLMs)

| Model | Best For | Rate Limit | Size | Notes | |-------|----------|------------|------|-------| | 2025 Models | | @cf/meta/llama-4-scout-17b-16e-instruct | Latest Llama, general purpose | 300/min | 17B | NEW 2025 | | @cf/openai/gpt-oss-120b | Largest open-source GPT | 300/min | 120B | NEW 2025 | | @cf/openai/gpt-oss-20b | Smaller open-source GPT | 300/min | 20B | NEW 2025 | | @cf/google/gemma-3-12b-it | 128K context, 140+ languages | 300/min | 12B | NEW 2025, vision | | @cf/mistralai/mistral-small-3.1-24b-instruct | Vision + tool calling | 300/min | 24B | NEW 2025 | | @cf/qwen/qwq-32b | Reasoning, complex tasks | 300/min | 32B | NEW 2025 | | @cf/qwen/qwen2.5-coder-32b-instruct | Coding specialist | 300/min | 32B | NEW 2025 | | @cf/qwen/qwen3-30b-a3b-fp8 | Fast quantized | 300/min | 30B | NEW 2025 | | @cf/ibm-granite/granite-4.0-h-micro | Small, efficient | 300/min | Micro | NEW 2025 | | Performance (2025) | | @cf/meta/llama-3.3-70b-instruct-fp8-fast | 2-4x faster (2025 update) | 300/min | 70B | Speculative decoding | | @cf/meta/llama-3.1-8b-instruct-fp8-fast | Fast 8B variant | 300/min | 8B | - | | Standard Models | | @cf/meta/llama-3.1-8b-instruct | General purpose | 300/min | 8B | - | | @cf/meta/llama-3.2-1b-instruct | Ultra-fast, simple tasks | 300/min | 1B | - | | @cf/deepseek-ai/deepseek-r1-distill-qwen-32b | Coding, technical | 300/min | 32B | - |

Text Embeddings (2x Faster - 2025)

| Model | Dimensions | Best For | Rate Limit | Notes | |-------|-----------|----------|------------|-------| | @cf/google/embeddinggemma-300m | 768 | Best-in-class RAG | 3000/min | NEW 2025 | | @cf/baai/bge-base-en-v1.5 | 768 | General RAG (2x faster) | 3000/min | pooling: "cls" recommended | | @cf/baai/bge-large-en-v1.5 | 1024 | High accuracy (2x faster) | 1500/min | pooling: "cls" recommended | | @cf/baai/bge-small-en-v1.5 | 384 | Fast, low storage (2x faster) | 3000/min | pooling: "cls" recommended | | @cf/qwen/qwen3-embedding-0.6b | 768 | Qwen embeddings | 3000/min | NEW 2025 |

CRITICAL (2025): BGE models now support pooling: "cls" parameter (recommended) but NOT backwards compatible with pooling: "mean" (default).

Image Generation

| Model | Best For | Rate Limit | Notes | |-------|----------|------------|-------| | @cf/black-forest-labs/flux-1-schnell | High quality, photorealistic | 720/min | ⚠️ See warnings below | | @cf/leonardo/lucid-origin | Leonardo AI style | 720/min | NEW 2025, requires numsteps | | @cf/leonardo/phoenix-1.0 | Leonardo AI variant | 720/min | NEW 2025, requires numsteps | | @cf/stabilityai/stable-diffusion-xl-base-1.0 | General purpose | 720/min | Requires num_steps |

⚠️ Common Image Generation Issues:

• Error 1000: Always include num_steps: 4 parameter (required despite docs suggesting optional)
• Error 3030 (NSFW filter): Single words like "hamburger" may trigger false positives - add descriptive context to prompts

// ✅ Correct pattern for image generation
const image = await env.AI.run('@cf/black-forest-labs/flux-1-schnell', {
  prompt: 'A photo of a delicious hamburger on a plate with fresh vegetables',
  num_steps: 4, // Required to avoid error 1000
});
// Descriptive context helps avoid NSFW false positives (error 3030)

Vision Models

| Model | Best For | Rate Limit | Notes | |-------|----------|------------|-------| | @cf/meta/llama-3.2-11b-vision-instruct | Image understanding | 720/min | - | | @cf/google/gemma-3-12b-it | Vision + text (128K context) | 300/min | NEW 2025 |

Audio Models (2025)

| Model | Type | Rate Limit | Notes | |-------|------|------------|-------| | @cf/deepgram/aura-2-en | Text-to-speech (English) | 720/min | NEW 2025 | | @cf/deepgram/aura-2-es | Text-to-speech (Spanish) | 720/min | NEW 2025 | | @cf/deepgram/nova-3 | Speech-to-text (+ WebSocket) | 720/min | NEW 2025 | | @cf/openai/whisper-large-v3-turbo | Speech-to-text (faster) | 720/min | NEW 2025 |

---

Common Patterns

RAG (Retrieval Augmented Generation)

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

// 2. Search Vectorize
const matches = await env.VECTORIZE.query(embeddings.data[0], { topK: 3 });
const context = matches.matches.map((m) => m.metadata.text).join('\n\n');

// 3. Generate 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,
});

---

Structured Output with Zod

import { z } from 'zod';

const Schema = z.object({ name: z.string(), items: z.array(z.string()) });

const response = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
  messages: [{
    role: 'user',
    content: `Generate JSON matching: ${JSON.stringify(Schema.shape)}`
  }],
});

const validated = Schema.parse(JSON.parse(response.response));

---

AI Gateway Integration

Provides caching, logging, cost tracking, and analytics for AI requests.

Basic Gateway Usage

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

// Access logs and send feedback
const gateway = env.AI.gateway('my-gateway');
await gateway.patchLog(env.AI.aiGatewayLogId, {
  feedback: { rating: 1, comment: 'Great response' },
});

Per-Request Cache Control (Advanced)

Override default cache behavior with HTTP headers for fine-grained control:

// Custom cache TTL (1 hour for expensive queries)
const response = await fetch(
  `https://gateway.ai.cloudflare.com/v1/${accountId}/${gatewayId}/workers-ai/@cf/meta/llama-3.1-8b-instruct`,
  {
    method: 'POST',
    headers: {
      'Authorization': `Bearer ${env.CLOUDFLARE_API_KEY}`,
      'Content-Type': 'application/json',
      'cf-aig-cache-ttl': '3600', // 1 hour in seconds (min: 60, max: 2592000)
    },
    body: JSON.stringify({
      messages: [{ role: 'user', content: prompt }],
    }),
  }
);

// Skip cache for real-time data
const response = await fetch(gatewayUrl, {
  headers: {
    'cf-aig-skip-cache': 'true', // Bypass cache entirely
  },
  // ...
});

// Check if response was cached
const cacheStatus = response.headers.get('cf-aig-cache-status'); // "HIT" or "MISS"

Available Cache Headers:

• cf-aig-cache-ttl: Set custom TTL in seconds (60s to 1 month)
• cf-aig-skip-cache: Bypass cache entirely ('true')
• cf-aig-cache-key: Custom cache key for granular control
• cf-aig-cache-status: Response header showing "HIT" or "MISS"

Benefits: Cost tracking, caching (reduces duplicate inference), logging, rate limiting, analytics, per-request cache customization.

---

Rate Limits & Pricing (Updated 2025)

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 | Im

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Kgeminic
• Source: Kgeminic/claude-skills-1
• License: MIT

Install and usage instructions live in the source repository linked above.