Cloudflare Clef is a new pair of open-weight AI models announced on October 1, 2026, built for one specific job: making decisions. Unlike ChatGPT or Gemini, Clef never writes paragraphs. You feed it some data and a few typed questions, and it answers with probabilities: how likely the answer is yes, which option from your list fits best, or where something lands on a scale you defined. It is the latest sign that the AI race is splitting into two tracks, the models that talk, and the models that decide.
What Is Cloudflare Clef?
Clef is what Cloudflare calls a decision model. It sits between a traditional classifier and a general-purpose large language model. A rigid classifier needs retraining every time your categories change. An LLM can reason about anything, but it answers every question by generating text, which you then have to parse, and that costs time and tokens.
A decision model takes a third path. You define the possible answers up front, ask the model a fixed set of typed questions about your input, and it returns a probability for each allowed answer. Your code reads the numbers and branches. No text to parse, no surprises.
Cloudflare launched two sizes. Clef is the 27 billion parameter flagship built for the most accurate calls. Clef-flash is the 9 billion parameter version built for answers in a few tens of milliseconds, aimed at latency-critical parts of an application where even a fraction of a second is too long. Both are built on a Qwen backbone and use a 64,000 token context window.
How Clef Works: Three Question Types
Clef supports three kinds of questions about any input you give it:
- Yes or no: returns the probability that the answer is yes. Is this support ticket urgent? Is this email phishing?
- Choice: picks one named option from your list, with a probability for each option and a confidence value. Which team should handle this request?
- Score: rates the input against an ordered scale you describe, returning a probability-weighted score. How severe is this incident?
On Workers AI, a single request can carry up to 64 questions and up to 4 images, because both models include a vision encoder. So you can pass a screenshot and ask whether it shows an error, which product is visible, and how serious the problem looks, all in one call.
Open Weights Under Apache 2.0
The detail that matters most is the license. Cloudflare published the actual model weights on Hugging Face under the Apache 2.0 license, which means developers can download them and run them on their own hardware instead of calling an API. That is a meaningful move from a company whose business is routing other people’s internet traffic: instead of metering access to the models, it gave the weights away.
Both models also run today on Cloudflare’s Workers AI platform, so you can call them through an API if you prefer. Pricing is set at $0.24 per million input tokens for Clef and $0.09 per million input tokens for Clef-flash, and output tokens are not billed at all, since the models return numbers rather than text.
The Benchmarks Cloudflare Is Citing
Cloudflare is positioning Clef directly against Jev, the decision model TypeSafe AI launched on September 15, 2026. In fact, Clef uses the same API format as Jev, so teams already built around Jev can switch by changing the endpoint and model name.
On the BANKING77 intent-classification benchmark, Cloudflare reports a macro-F1 score of 94.20 for Clef against 79.74 for Jev. On function-call accuracy (BFCL), Clef scored 98.47 against 95.75. On speed, Clef-flash posted a median latency of 38.8 milliseconds where Jev reported 524.1 milliseconds. Both Clef variants also handle images, which Jev does not currently support. As with all vendor-reported benchmarks, treat these as a starting point rather than a final verdict until independent testing confirms them.
Why Decision Models Matter for AI Agents
Most AI agents still work the same wasteful way under the hood. The agent asks a large language model a yes-or-no question, and the model generates a full sentence to answer it, burning time and tokens just to say yes. For agents that need to make hundreds of small decisions, sorting tickets, routing requests, approving steps, that overhead adds up fast.
Decision models are built for exactly this layer. A support workflow could pass in a customer message and ask whether it is urgent, which team should own it, and how severe the impact is, then route the ticket or escalate to a human based on the probabilities. Cloudflare’s own Threat Intelligence team tested Clef on domain classification: given a website, it returned category probabilities like 95% fashion, 85% ecommerce, under 1% phishing, completing the whole fetch, render, and classify pipeline in 2.2 seconds against 4.7 seconds for a general-purpose model.
This category is suddenly crowded. TypeSafe launched Jev in mid-September, OpenAI announced its Decisions API in limited preview on September 29, and Amazon shipped Strands Decider 2B the same day Cloudflare announced Clef. Cloudflare is the only one of the three giving away the weights, which may be what decides the category.
Reinforcement Learning Fine-Tuning
Alongside the models, Cloudflare announced a reinforcement learning service that lets customers capture their own request data, fine-tune Clef on it, and redeploy their own variants. It starts as a hands-on engagement with Cloudflare’s engineering team, with a self-serve version planned later. Cloudflare states it does not read, store, or train on customer requests unless they opt into this product.
Who Should Pay Attention
If you build anything with AI agents, Clef is worth a look. Support automation, operations triage, content moderation queues, fraud review, any workflow where software needs to make fast, structured judgment calls. If you are simply choosing a chatbot for your small business, a decision model is not what you need. Our Claude vs ChatGPT comparison covers which conversational AI is worth paying for in 2026.
For developers already experimenting with AI tooling, the rise of decision models is part of the same shift we covered in our comparison of AI coding assistants and the big model launches like Gemini 4 Argon: AI is moving from impressive demos into production infrastructure, and the winners will be the tools that are fast, cheap, and predictable.
FAQ
What is Cloudflare Clef?
Clef is a pair of open-weight AI decision models released by Cloudflare on October 1, 2026. Instead of generating text like ChatGPT, they answer typed questions with probabilities, which makes them fast and cheap for automated workflows.
How is Clef different from ChatGPT or Gemini?
ChatGPT and Gemini are conversational models that generate free-form text. Clef never writes text. It takes your data and a fixed set of questions, then returns structured probabilities your code can act on directly, without parsing anything.
Is Cloudflare Clef free to use?
The model weights are free: they are published on Hugging Face under the Apache 2.0 license, so you can download and self-host them. Through Cloudflare’s Workers AI API, pricing is $0.24 per million input tokens for Clef and $0.09 per million for Clef-flash, with no charge for outputs.
How fast is Clef?
Cloudflare reports a median latency of 38.8 milliseconds for Clef-flash, versus 524.1 milliseconds for TypeSafe’s rival Jev model. Independent verification is still pending.
What is a decision model?
A decision model sits between rigid classifiers and general-purpose LLMs. It answers a fixed set of typed questions about an input with calibrated probabilities, without retraining when your categories change, and without the latency and cost of generating text.