navhunt.
Clef AI logo

Clef AI

Clef AI is the pair of decision models Cloudflare shipped on 1 October 2026. The larger model is 27B and is hosted as @cf/cloudflare/clef. Clef-flash is 9B and is hosted as @cf/cloudflare/clef-flash. You send a state and up to 64 typed questions. Clef AI returns a probability for each allowed answer, not an essay and not a tool-call transcript.

Visit Clef AISee pricing

Official website · developers.cloudflare.com

A NavHunt schematic of the documented Clef AI request. It is not a product screenshot and not a benchmark chart. Source ↗

Announced
1 October 2026
Hosted sizes
27B and 9B
Hosted context
65,536 tokens
Questions
1 to 64 per request
Weight license
Apache-2.0
Plan ladder
No Clef-specific tier confirmed

Is Clef AI right for you?

Clef AI fits a closed choice: which queue, which severity, or whether yes is likely enough to continue. It is the wrong tool for new wording or a decision a person must own. Weights do not make hosted inference free, and the local helper is not the hosted request.

A good fit for

  • Routing a record into buckets you have already named.
  • Choosing the 27B or 9B hosted model from published rates and limits.
  • Seeing where the Workers AI schema and the weight-card helper differ.

Know the trade-offs

  • A probability from Clef AI is a model score, not proof the decision is correct or calibrated.
  • The pricing table lists input rates only, while the response still has an output-token field.
  • Fine-tuning has no public date or price.

What a Clef AI call is allowed to return

NavHunt explanatory diagram of a Clef AI request, with state and typed questions returning probabilities, beside the local weight helper
A NavHunt schematic of the documented Clef AI request. It is not a product screenshot and not a benchmark chart. Source

Clef AI decides inside a schema

The 1 October changelog describes Clef AI as a decision model in Jev's family. It returns a probability for every allowed answer, not a paragraph. A Text Generation label, or the card tag image-text-to-text, can show a chat box. That box is not the schema.

The hosted body is model, state, and 1 to 64 questions. model is clef or clef-flash and must match the id you called. state is text or JSON. Ids may use letters, digits, underscore, dot, and hyphen.

Noul is the probability of yes, from 0 to 1, with no confidence field. Choice has 2 to 255 options and returns the top option, probabilities, and confidence. Score has 2 to 10 levels and returns a weighted score, a legend, probabilities, and confidence. Those probabilities sum to 1. Confidence is derived from them.

Official documentation

The two Clef AI models are not the same request

Cloudflare points the 27B model at precision and Clef-flash, the 9B model, at latency. Both hosted pages list 65,536 tokens. The cards name frozen Qwen/Qwen3.8-27B and Qwen/Qwen3.5-9B bases plus a joint scoring head. Rank-256 adapters and a Brier loss are training details, not a promise that a live score is calibrated.

Call env.AI.run or POST /ai/run. Instructions are required. Up to four embedded PNG, JPEG, or WebP images are allowed, not remote URLs, at 4 MiB and 16 megapixels each, 8 MiB decoded in total, inside a 13 MiB body.

The model page mentions video, but the hosted schema has no videos field. The card's local helper adds video frames, optional instructions, and a 16,384-token default. A chat sample matches neither body.

Official documentation

A queue decision shows the Clef AI shape

Cloudflare's support example asks Clef AI about urgency, team, and severity in one call. A score should not close an account alone. Clef AI does not fetch an image URL. The launch post by Michelle Chen, Alex Reneau, and Kevin Flansburg includes a vendor table against Jev. This page does not rank those models.

Official documentation

Ways to use Clef AI

Queue a support note

Ask Clef AI whether a message is urgent, which named team owns it, and how severe it is.

Look at a screenshot before a tool runs

Before a tool runs, ask Clef AI one yes-or-no question and attach up to four embedded images. A page URL is not accepted. The probability is not permission.

Score one record in several ways

One JSON refund can carry several questions in a single Clef AI call, up to 64. You still write every allowed answer. Missing criteria are not invented.

How to get started with Clef AI

  1. 1

    Decide whether the output is a choice

    If you need a paragraph, Clef AI is the wrong call. If every answer is listed, use noul, choice, or score.

  2. 2

    Pick a hosted size or the weights

    Use a Workers AI id for the 65,536-token window, or the weight repo only if you will run the helper yourself.

  3. 3

    Send the documented body

    Pass model, state, and questions. Embed PNG, JPEG, or WebP bytes only. Ignore a generic chat snippet.

  4. 4

    Treat the numbers as model scores

    Compare the probability with a threshold you chose. Keep a person on safety, money, and access decisions.

Clef AI rates that are published, and bills that are not

Check live pricing

Clef AI has no confirmed Free, Pro, or Enterprise ladder of its own. The figures are unit rates from the Workers AI table updated 1 October 2026 and re-read on 7 October 2026. NavHunt did not buy a subscription.

27B hosted input

Hosted calls to @cf/cloudflare/clef

$0.24

per million input tokens

  • Also listed as 21,818 Neurons per million input tokens
  • 65,536-token context on the model page
  • No separate output price on this row

9B hosted input

Hosted calls to @cf/cloudflare/clef-flash

$0.09

per million input tokens

  • Also listed as 8,182 Neurons per million input tokens
  • Same 65,536-token hosted window
  • No separate output price on this row

Weight download

Run the card helper on your own hardware

$0

Apache-2.0 license fee

  • The license fee is not a GPU or power budget
  • Local encode default is 16,384 tokens
  • Fine-tuning had no public price on the pages read

Workers AI also lists $0.011 per 1,000 Neurons. Times 21,818 and 8,182, that matches the token prices, so the columns are one charge. Free and Paid plans share 10,000 Neurons a day. Beyond that, Workers Paid is required. NavHunt's division is about 458,000 input tokens for the 27B model and 1.22 million for Clef-flash. The pool is not reserved for Clef AI, and no call proved these models draw it down. output_tokens is still in the schema. A missing output price does not prove output is free. The blog's no-training claim was not audited. Fine-tuning has no public date or price.

Clef AI: common questions

Is Clef AI just a chatbot with another name?

No. Clef AI returns probabilities for questions you wrote. Noul is one number, choice names the top option, and score is a weighted level rather than a paragraph. A hub chat box follows the generic pipeline tag. The card says to load joint_schema_model and score a record.

Source
Which Clef AI sample is the real interface?

Hosted calls use the model page and the two schemas: model, state, questions, and optional images. Weights use load_release_model and the systemone helper on the handwritten card. An automatic chat sample is neither. The local helper documents videos and optional instructions. The hosted schema read on 7 October 2026 does not.

Source
How should the two Clef AI models be chosen?

Cloudflare aims the 27B model at precision and Clef-flash at latency. Both hosted pages list 65,536 tokens and the same question types, at $0.24 and $0.09 per million input tokens. The cards name Qwen/Qwen3.8-27B and Qwen/Qwen3.5-9B, and both base cards mark Apache-2.0. NavHunt did not measure speed or copy a finer parameter count.

Source
Does a high Clef AI probability mean the decision is safe?

No. Values sit between 0 and 1, and choice and score probabilities sum to 1. Confidence is derived from those probabilities, not from your own outcomes. A training loss does not make 0.95 a measured hit rate. Do not let Clef AI replace a person on a hard-to-undo decision.

Source
How does Clef AI relate to Jev, without naming a winner?

Change the endpoint and model name, and keep state plus the three question types. Hosted images are an added field. The local helper can map that body on your machine. The blog calls Jev text-only at 32K and prints scores. NavHunt did not check those limits or rerun them, so Clef AI is described here only at the API layer.

Source
Does Apache-2.0 make hosted Clef AI free?

No. The weight file is Apache-2.0, following the named Qwen base. Workers AI still lists input rates, and a Neuron pool is platform policy, not part of that license. Hardware you run is also outside it. A zero license fee is not a zero bill.

Source
Can one hosted Clef AI request send a video or an image URL?

Not in the schema read on 7 October 2026. Hosted images are embedded PNG, JPEG, or WebP, at most four, 4 MiB and 16 megapixels each, 8 MiB decoded in total, inside a 13 MiB body. Remote URLs are refused. Video frames are described only for the local helper. The model page mentions video in prose, so the page and the schema disagree until a hosted videos field exists. Do not send a video URL to the REST route.

Source

Sources & last checked

Oct 7, 2026 · NavHunt editorial

Changelog, blog, model pages, schemas, pricing, both cards, the Apache file, and Qwen license metadata were read on 7 October 2026. No API call or benchmark rerun. Canonical URL: https://navhunt.com/products/clef. The image is an original NavHunt explanatory diagram.

Use the Clef AI schema, not a chat sample

Start from the Workers AI model page or the handwritten weight card. NavHunt does not proxy this model.

Explore Clef AI