Cloudflare Workers AI Models Directory

Cloudflare Workers AI lets you run AI models on Cloudflare's global network, powered by serverless GPUs. You do not rent a GPU, install anything, or manage a server. You send a request from a Worker, a Pages Function, or the REST API, and you get the result back. The models are open-weight models — Llama, Qwen, Gemma, FLUX, Whisper, and many more — ready to use from your own code.

This page is a simple directory of every model in the Workers AI catalog. Use it to find the right model for your job: a chatbot, an image generator, a speech transcriber, a translator, or a search engine. Each model has its official name, its model ID (the exact name you use in code), and one plain-English line about what it does.

Quick answer: Cloudflare's official Workers AI catalog lists 73 models (checked 12 November 2026). The free plan gives you 10,000 Neurons per day at no charge, and the limit resets every day at 00:00 UTC. If you go over the limit, further requests fail until the next reset. This page is a reference directory only — there is no live playground here, and nothing on this page calls the Cloudflare API.

Last checked: 12 November 2026. Cloudflare adds and removes models often, so check the official Workers AI catalog for the newest list.

What is Cloudflare Workers AI?

Workers AI is Cloudflare's way of running machine learning models at the edge — inside its network of data centres around the world, close to your users. Instead of sending your request to one far-away GPU server, the work runs at the Cloudflare location nearest to you.

You use it in three ways. First, with a Workers Binding: your Cloudflare Worker calls the model directly in code. Second, with the REST API: any app can send an HTTP request to Cloudflare and get a result back. Third, with OpenAI-compatible endpoints: if your code already talks to OpenAI's API, you can point it at Workers AI with almost no changes.

The catalog covers the most common AI jobs: writing and chatting with text, making images from text, turning speech into text, turning text into speech, translating between languages, searching by meaning, checking if text is positive or negative, and understanding what is in a picture. What it does not have is frontier closed models — you will not find GPT-4o or Claude in this catalog. Those live in Cloudflare's separate unified AI catalog, which lists third-party models served through AI Gateway.

The free limit, in plain English

Workers AI charges in Neurons. A Neuron is Cloudflare's unit for measuring GPU work: different models cost a different number of Neurons for the same request, because bigger models need more compute. You do not pay per request — you pay in Neurons, and the price is $0.011 per 1,000 Neurons once you go past the free amount.

The free amount is 10,000 Neurons per day, on both the Free and Paid Workers plans. All limits reset daily at 00:00 UTC. If you use up your 10,000 Neurons, further requests fail with an error until the next reset — on the Free plan there is no automatic charge, it simply stops. On the Paid plan, usage above the free 10,000 Neurons is charged at $0.011 per 1,000 Neurons.

What does 10,000 Neurons buy? It depends on the model. A few rough examples from Cloudflare's own pricing table:

  • About 500 short chats a day on Llama 3.1 8B — a short question and answer costs roughly 20 Neurons.
  • About 240 minutes of audio transcription a day with Whisper, which costs 41.14 Neurons per audio minute.
  • About 9 million tokens of embeddings a day with BGE-M3, which costs 1,075 Neurons per million input tokens.

Small models are cheap; big frontier models are expensive. That is why the same 10,000 Neurons can last all day for a small chatbot but only a few minutes for a giant reasoning model.

Two more things to know before you pick a model:

  • 8 models need a paid billing method. These frontier models — Kimi K2.6, Kimi K2.7 Code, GLM-5.2, GLM-5.3, GLM-5.3 Flash, DeepSeek V4 Flash, DeepSeek V4.1 Flash, and DeepSeek V4 Pro — need the Workers Paid plan or prepaid AI Gateway credits. They are marked in the list below.
  • 2 models need an access form. Apertus 1.5 8B and EuroLLM 9B ask you to fill in a form before you can use them. They are marked below too.

There are also speed limits per task. Text generation allows 300 requests per minute, text embeddings 3,000 per minute, image classification 3,000 per minute, and speech recognition 720 per minute. Models that need paid billing have lower limits: 20 requests per minute on standard billing, or 50 with prepaid AI Gateway credits.

A note about the number 276

You may have seen the number 276 mentioned for Cloudflare's AI models. That number does not match the official Workers AI catalog, which lists 73 models (checked 12 November 2026, confirmed on both the catalog page and Cloudflare's own documentation index). We have listed all 73 here — nothing added, nothing padded.

Bigger numbers come from Cloudflare's unified AI catalog, a separate and much larger list that includes third-party models from other companies (for example Claude, GPT, and video models) served through AI Gateway. Those are not Workers AI models, so they are not in this directory. If a list claims hundreds of models, it is almost certainly counting the unified catalog, not Workers AI.

Most popular models — start here

If you are new to Workers AI, these are the best-known models in the catalog — the ones most developers try first:

  • Llama 3.3 70B (@cf/meta/llama-3.3-70b-instruct-fp8-fast) — the big, smart all-rounder for chat, coding, and agents.
  • Llama 3.1 8B (@cf/meta/llama-3.1-8b-instruct-fp8) — the cheap, fast default for everyday chatbots and small apps.
  • Llama 4 Scout (@cf/meta/llama-4-scout-17b-16e-instruct) — the Llama that can also read images.
  • gpt-oss-120b (@cf/openai/gpt-oss-120b) — OpenAI's open-weight model for hard reasoning and agentic work.
  • DeepSeek-R1-Distill-Qwen-32B (@cf/deepseek-ai/deepseek-r1-distill-qwen-32b) — the thinking model for maths and hard problems.
  • FLUX.1 [schnell] (@cf/black-forest-labs/flux-1-schnell) — the fastest way to turn text into images.
  • Whisper (@cf/openai/whisper) — the standard choice for turning speech into text.
  • BGE-M3 (@cf/baai/bge-m3) — the multilingual embedding model for search and RAG.

How to use this directory

Every model has an ID that looks like this: @cf/meta/llama-3.3-70b-instruct-fp8-fast. The @cf part means it is hosted by Cloudflare, the middle part is the author (Meta, OpenAI, Google, and so on), and the last part is the model name. Copy the full ID into your Worker code or API request exactly as shown.

Click any model name to open its official Cloudflare documentation page, with its full details, inputs, outputs, and example code. The tables below are grouped by task, so you can compare all the models that do the same job.

Text Generation — 39 models

Chatbots, writing, coding, reasoning, and agents. This is the biggest group in the catalog.

ModelModel IDWhat it does
apertus-v1.5-8b †@cf/swiss-ai/apertus-v1.5-8bOpen 8B model for many languages, up to 262K context. Needs an access form.
clairic@cf/cloudflare/clairicPulls out numbers, estimates, and forecasts from your input, with confidence scores. In preview.
clef@cf/cloudflare/clef27B decision model: reads text, JSON, images, or video and answers your questions with probabilities.
clef-flash@cf/cloudflare/clef-flashFaster 9B version of Clef for quick decisions.
clef-omni@cf/cloudflare/clef-omniDecision model that also reads audio; 30B mixture-of-experts, 3B active.
deepseek-r1-distill-qwen-32b@cf/deepseek-ai/deepseek-r1-distill-qwen-32bReasoning model distilled from DeepSeek-R1; thinks step by step before answering.
deepseek-v4-flash-0731 ‡@cf/deepseek-ai/deepseek-v4-flash-0731Fast frontier model with strong agentic skills. Needs paid billing.
deepseek-v4-pro-0813 ‡@cf/deepseek-ai/deepseek-v4-pro-0813High-capability reasoning model with a 1M-token context. Needs paid billing.
deepseek-v4.1-flash-1017 ‡@cf/deepseek-ai/deepseek-v4.1-flash-1017Multimodal model (552B backbone) that reads images and text. Needs paid billing.
eurollm-9b-it †@cf/utter-project/eurollm-9b-it9B model trained on European languages. Needs an access form.
gemma-2b-it-lora@cf/google/gemma-2b-it-loraSmall Gemma base model for running your own LoRA fine-tunes. Beta.
gemma-4-26b-a4b-it@cf/google/gemma-4-26b-a4b-itGoogle's smartest open model family, built from Gemini 3 research.
gemma-7b-it-lora@cf/google/gemma-7b-it-loraGemma 7B base model for running your own LoRA fine-tunes. Beta.
gemma-sea-lion-v4-27b-it@cf/aisingapore/gemma-sea-lion-v4-27b-itLanguage model tuned for Southeast Asian languages.
glm-4.7-flash@cf/zai-org/glm-4.7-flashFast multilingual chat model, 131K context, 100+ languages.
glm-5.2 ‡@cf/zai-org/glm-5.2Flagship agentic coding model from Z.ai. Needs paid billing.
glm-5.3 ‡@cf/zai-org/glm-5.3Flagship agentic coding model with a 1M-token context. Needs paid billing.
glm-5.3-flash ‡@cf/zai-org/glm-5.3-flashFirst multimodal GLM-5 model, fast and cheap for its size. Needs paid billing.
gpt-oss-120b@cf/openai/gpt-oss-120bOpenAI's open-weight reasoning model for production, general-purpose work.
gpt-oss-20b@cf/openai/gpt-oss-20bSmaller, lower-latency version of OpenAI's open-weight reasoning model.
granite-4.0-h-micro@cf/ibm-granite/granite-4.0-h-microIBM model built for agents, function calling, and RAG.
kimi-k2.6 ‡@cf/moonshotai/kimi-k2.6Frontier 1T-parameter model, 262K context, tool calling and vision. Needs paid billing.
kimi-k2.7-code ‡@cf/moonshotai/kimi-k2.7-codeFrontier 1T-parameter model tuned for coding and agents. Needs paid billing.
llama-2-7b-chat-hf-lora@cf/meta-llama/llama-2-7b-chat-hf-loraLlama 2 chat base model for running your own LoRA fine-tunes. Beta.
llama-3.1-8b-instruct-fp8@cf/meta/llama-3.1-8b-instruct-fp8Fast, cheap 8B Llama 3.1 for chat and general text.
llama-3.2-11b-vision-instruct@cf/meta/llama-3.2-11b-vision-instructLlama that can see: answers questions about images and describes them.
llama-3.2-1b-instruct@cf/meta/llama-3.2-1b-instructTiny 1B Llama for simple chat and summarising on a budget.
llama-3.2-3b-instruct@cf/meta/llama-3.2-3b-instructSmall 3B Llama for multilingual chat and retrieval tasks.
llama-3.3-70b-instruct-fp8-fast@cf/meta/llama-3.3-70b-instruct-fp8-fastBig 70B Llama, optimised to run faster.
llama-4-scout-17b-16e-instruct@cf/meta/llama-4-scout-17b-16e-instructMultimodal 17B Llama (16 experts) for text and image understanding.
llama-guard-3-8b@cf/meta/llama-guard-3-8bSafety checker: reads a prompt or a reply and says whether it is safe or unsafe.
mistral-7b-instruct-v0.2-lora@cf/mistral/mistral-7b-instruct-v0.2-loraMistral 7B instruct base model for running your own LoRA fine-tunes.
mistral-small-3.1-24b-instruct@cf/mistralai/mistral-small-3.1-24b-instruct24B Mistral with vision and 128K context.
nemotron-3-120b-a12b@cf/nvidia/nemotron-3-120b-a12bNVIDIA's big model for multi-agent and agentic systems.
phonellm-mini-alpha-2@cf/daily-co/phonellm-mini-alpha-2Language model tuned for voice agents and phone calls.
qwen2.5-coder-32b-instruct@cf/qwen/qwen2.5-coder-32b-instructQwen model built for writing and explaining code.
qwen3-30b-a3b-fp8@cf/qwen/qwen3-30b-a3b-fp8Qwen3 mixture-of-experts model for reasoning, agents, and multilingual chat.
qwen3.8-27b@cf/qwen/qwen3.8-27b27B Qwen model for vision, general chat, and agentic work.
qwq-32b@cf/qwen/qwq-32bQwen's reasoning model: thinks through hard problems before answering.

† Needs an access form before you can use it. ‡ Needs a paid billing method (Workers Paid plan or prepaid AI Gateway credits).

Text-to-Image — 10 models

Turn a text prompt into a picture. FLUX models are the newest and most capable here; Stable Diffusion models are the classic open choice.

ModelModel IDWhat it does
dreamshaper-8-lcm@cf/lykon/dreamshaper-8-lcmStable Diffusion tuned for photorealistic images.
flux-1-schnell@cf/black-forest-labs/flux-1-schnellFast 12B image model; turns text into images.
flux-2-dev@cf/black-forest-labs/flux-2-devRealistic, detailed images with multi-reference support.
flux-2-klein-4b@cf/black-forest-labs/flux-2-klein-4bUltra-fast small image model for generation and editing.
flux-2-klein-9b@cf/black-forest-labs/flux-2-klein-9b9B image model for generation and multi-reference editing.
lucid-origin@cf/leonardo/lucid-originLeonardo's most prompt-faithful image model, good at text inside images.
phoenix-1.0@cf/leonardo/phoenix-1.0Leonardo image model with strong prompt following and clear text.
stable-diffusion-v1-5-inpainting@cf/runwayml/stable-diffusion-v1-5-inpaintingFills in or replaces part of a picture using a mask. Beta.
stable-diffusion-xl-base-1.0@cf/stabilityai/stable-diffusion-xl-base-1.0Creates and edits images from text prompts. Beta.
stable-diffusion-xl-lightning@cf/bytedance/stable-diffusion-xl-lightningVery fast text-to-image model; high-quality 1024px images in a few steps. Beta.

Text Embeddings — 7 models

Turn text into numbers (vectors) so a computer can compare meaning — the foundation of search, recommendations, and RAG chatbots.

ModelModel IDWhat it does
bge-base-en-v1.5@cf/baai/bge-base-en-v1.5Turns text into a 768-number vector for search and matching.
bge-large-en-v1.5@cf/baai/bge-large-en-v1.5Turns text into a 1024-number vector; the bigger, more accurate BGE.
bge-m3@cf/baai/bge-m3Multilingual embeddings that work in many languages and at different text lengths.
bge-small-en-v1.5@cf/baai/bge-small-en-v1.5Turns text into a 384-number vector; the smallest, fastest BGE.
embeddinggemma-300m@cf/google/embeddinggemma-300mGoogle's small embedding model, trained in 100+ languages. Beta.
plamo-embedding-1b@cf/pfnet/plamo-embedding-1bEmbedding model built for Japanese text.
qwen3-embedding-0.6b@cf/qwen/qwen3-embedding-0.6bQwen's embedding model for search and ranking.

Automatic Speech Recognition — 5 models

Turn spoken audio into written text — for transcription, subtitles, and voice apps.

ModelModel IDWhat it does
flux@cf/deepgram/fluxSpeech recognition built for live voice agents and conversations.
nova-3@cf/deepgram/nova-3Deepgram's speech-to-text model for transcribing audio.
whisper@cf/openai/whisperTurns speech into text in many languages; also translates speech and detects the language.
whisper-large-v3-turbo@cf/openai/whisper-large-v3-turboFaster version of Whisper Large for transcription and speech translation.
whisper-tiny-en@cf/openai/whisper-tiny-enTiny English-only Whisper for quick transcription. Beta.

Text-to-Speech — 4 models

Turn written text into spoken audio — for voiceovers, audiobooks, and voice assistants.

ModelModel IDWhat it does
aura-1@cf/deepgram/aura-1Turns text into natural speech with pacing and expression.
aura-2-en@cf/deepgram/aura-2-enNewer English text-to-speech voice with natural pacing and fillers.
aura-2-es@cf/deepgram/aura-2-esNewer Spanish text-to-speech voice with natural pacing and fillers.
melotts@cf/myshell-ai/melottsHigh-quality text-to-speech in many languages.

Translation — 2 models

Translate text from one language to another.

ModelModel IDWhat it does
indictrans2-en-indic-1B@cf/ai4bharat/indictrans2-en-indic-1BTranslates between English and 22 Indian languages.
m2m100-1.2b@cf/meta/m2m100-1.2bTranslates directly between 100 languages, with no English needed in the middle.

Text Classification — 2 models

Sort text into categories, or score how well two texts match.

ModelModel IDWhat it does
bge-reranker-base@cf/baai/bge-reranker-baseScores how well a document matches a question, so you can rank search results.
distilbert-sst-2-int8@cf/huggingface/distilbert-sst-2-int8Small model that tells you if a sentence is positive or negative.

Image-to-Text — 2 models

Models that look at a picture and describe it in words, or answer questions about it.

ModelModel IDWhat it does
llava-1.5-7b-hf@cf/llava-hf/llava-1.5-7b-hfChatbot that looks at a picture and talks about it. Beta.
moondream3.1-9B-A2B@cf/moondream/moondream3.1-9B-A2BFast vision model for describing images, finding objects, and reading text in pictures.

Image Classification — 1 model

Identify what is in a picture.

ModelModel IDWhat it does
resnet-50@cf/microsoft/resnet-50Classic image classifier trained on 1M+ photos; tells you what is in a picture.

Voice Turn Detection — 1 model

Detect when a speaker has finished talking, so a voice agent knows when to reply. Cloudflare lists this model under the task name "Dumb Pipe".

ModelModel IDWhat it does
smart-turn-v2@cf/pipecat-ai/smart-turn-v2Listens to a conversation and detects when a speaker has finished, so a voice agent knows when to reply.

Frequently Asked Questions

How many models does Workers AI have?

73 models, checked on 12 November 2026 from Cloudflare's official catalog. The catalog page itself says "We found 73 models". If you see a bigger number like 276, that counts Cloudflare's separate unified AI catalog, which includes third-party models — not the Workers AI catalog.

Is Workers AI really free?

Yes, up to a point. Every account gets 10,000 Neurons per day free, on both the Free and Paid Workers plans. When the free amount runs out, requests fail until the next daily reset at 00:00 UTC — on the Free plan you are not charged. On the Paid plan, usage above 10,000 Neurons per day costs $0.011 per 1,000 Neurons.

What is a Neuron?

A Neuron is Cloudflare's way of measuring the GPU work a request needs. It is the unit Workers AI bills in. Different models cost a different number of Neurons: small models cost few Neurons, big models cost many. Cloudflare publishes the exact Neuron rate for each model on its pricing page.

Can I try these models on this page?

No. This is a reference directory only — there is no live playground here, and this page does not call the Cloudflare API. To try a model, you need a Cloudflare account and a Worker or an API call. The free 10,000 Neurons per day is enough for testing and small projects.

Which model should I start with?

For a chatbot, start with Llama 3.1 8B (@cf/meta/llama-3.1-8b-instruct-fp8) — it is cheap and fast. For images, start with FLUX.1 [schnell]. For turning speech into text, start with Whisper. For search or a RAG chatbot, start with BGE-M3 for embeddings. Move to bigger models only when the small ones are not good enough for your job.

Why do some models say they need paid billing?

Eight frontier models — Kimi K2.6, Kimi K2.7 Code, GLM-5.2, GLM-5.3, GLM-5.3 Flash, DeepSeek V4 Flash, DeepSeek V4.1 Flash, and DeepSeek V4 Pro — need a paid billing method (the Workers Paid plan or prepaid AI Gateway credits). They are the biggest and most expensive models to run, so Cloudflare does not include them in the free allowance.

Are these models the same as ChatGPT or Claude?

No. Workers AI runs open-weight models — models anyone can download and inspect, like Llama, Qwen, and Gemma. You will not find ChatGPT, GPT-4o, or Claude in this catalog. Those models are available through Cloudflare's separate unified AI catalog via AI Gateway, as third-party models.

Do the models change over time?

Yes. Cloudflare adds new models and removes old ones regularly — the catalog page was last updated on 12 August 2026, and the pricing page on 20 October 2026. If a model ID in your code stops working, check the official catalog: it may have been retired or replaced by a newer version.

Source: Cloudflare's official Workers AI documentation. Model list, descriptions, and limits checked on 12 November 2026. See the official catalog at developers.cloudflare.com/workers-ai/models/.