Image creation
Cloudflare Clef
Cloudflare’s open-source decision models: ask questions, get probability-scored answers your code can act on.
Product insights
Cloudflare Clef at a glance
Value proposition
Reliable, machine-readable decisions with confidence scores, faster and cheaper than asking a chat LLM.
Problem solved
Chat LLMs answer in free text, so turning them into dependable yes/no or routing decisions takes fragile parsing and prompt tricks.
Audience
Developers building AI agents, support automation, moderation, or security triage on Cloudflare or any stack.
Pricing and access
Pay as you go on Workers AI (about $0.24 per million input tokens). The weights are free on Hugging Face under Apache 2.0.
Market
- AI decision and classification models
- AI agent infrastructure
Tech stack
- Clef on Qwen3.8-27B, Clef-flash on Qwen3.5-9B
- 64K context window with image input
- Workers AI: @cf/cloudflare/clef
01Why it is here
What is Cloudflare Clef?
Cloudflare Clef is a family of two “decision models”. You send a state (text or JSON, plus up to four images) and up to 64 questions. Each question is a yes/no check, a choice between named options, or a score on an ordered scale, and Clef returns a probability for every answer. Clef is the higher-accuracy model (about 209 ms median latency), while Clef-flash targets latency-critical calls (about 39 ms). Cloudflare reports leading results on intent benchmarks such as BANKING77 and CLINC150, and Clef-flash scores 98.76 on the BFCL tool-calling benchmark.
02Workflow
How to use Cloudflare Clef
- 01
Open Workers AI
Use a Cloudflare account and call @cf/cloudflare/clef or @cf/cloudflare/clef-flash from a Worker or the REST API.
- 02
Describe the state and questions
Pass the ticket, order, or page as state, then define questions as noul (yes/no), choice, or score.
- 03
Act on the probabilities
Read the answers and their probabilities, then route, escalate, or defer to a human when confidence is low.
03Capabilities
Key features of Cloudflare Clef
- Yes/no, choice, and score questions
- Probability for every answer
- Up to 64 questions per call
- Image input (up to 4 images)
- Open weights under Apache 2.0
04Use cases
What is Cloudflare Clef best for?
- Support ticket triage and routing
- Agent next-step and escalation decisions
- Moderation and security classification
05Questions
FAQ about Cloudflare Clef
What is Cloudflare Clef?
Clef is Cloudflare’s first open-source decision model family. It answers structured questions with probabilities instead of free text.
What is the difference between Clef and Clef-flash?
Clef is the more accurate model with vision. Clef-flash is smaller and much faster for latency-critical decisions.
Is Cloudflare Clef free?
The weights are free on Hugging Face. Running it on Workers AI is billed per token, at about $0.24 per million input tokens.
Official website snapshot
developers.cloudflare.com
Clef and Clef-flash are open-source (Apache 2.0) decision models from Cloudflare, launched on Workers AI on October 1, 2026. Instead of writing a paragraph, they return typed answers with probabilities: yes/no, pick one option, or a score. Agents and apps use them to route tickets, escalate risky cases, or decide the next step.
