the short answer
Cloudflare launched Clef and Clef-flash on Workers AI on October 1, 2026. It says the interface accepts Jev-style requests, and it publishes Clef weights and visual input support. Jev remains TypeSafe’s separate hosted model. Cloudflare’s own benchmark table has wins for each model on different tasks. Reuse your question definitions if the client passes a contract test, then measure accuracy and fit new thresholds on your labels.
- Launch
- Cloudflare announced Clef and Clef-flash October 1, 2026
- Interface
- Cloudflare claims Jev API compatibility; Workers AI documents state and typed questions
- Deployment
- Clef on Workers AI with Apache 2.0 open weights; Jev through TypeSafe and providers
- Modality
- Cloudflare documents Clef vision; TypeSafe documents Jev as text-only
- Clef price
- Workers AI lists $0.24 per million input tokens for Clef, checked October 2, 2026
The Name on the Route Matters
Cloudflare hosts both TypeSafe Jev and its own Clef family. The Clef model page shows state, Noul, Choice and Score questions on the @cf/cloudflare/clef route. A familiar request shape makes a trial easier to wire up. It says nothing about whether the two models assign the same probabilities.
As of October 2, Cloudflare listed Clef as a 27B multimodal model with a 65,536-token context window and Apache 2.0 weights. The launch post also introduced Clef-flash and assisted fine-tuning, with self-serve training planned for later. Check the live model listing for current availability before committing to either route.
Same Decision, Two Model Families
Cloudflare has Workers AI and HTTP request examples for the actual call. Here, the customer cannot sign in to download an invoice, so account access owns the first blocker. Nothing in the message supports a duplicate-charge finding. Use the current Cloudflare schema when you implement this fixture. The table records the question and expected label.
| Field | Jev or Clef experiment |
|---|---|
| State | Customer message: “I cannot sign in to download my invoice.” |
| Choice question | Which team resolves the first blocker: account access, billing, technical, or needs review? |
| Noul question | Does the message report a completed duplicate charge? |
| Application action | Route the ticket; require independent identity checks before any account action |
| Record | Provider, model ID, question version, full distributions, latency, failure and later human label |
Read Cloudflare’s Benchmark as a Vendor Result
Clef leads the aggregate in Cloudflare’s Jev Decision Index. Look one row deeper and Jev leads on When2Call and BRIGHT. Cloudflare ran these tests, so a support-routing or policy team still needs its own labels before choosing a model.
The latency figure belongs to Cloudflare’s test setup too. Region, batching, state length and retries can move it. Record those details in your own run, along with failed attempts, using the benchmark protocol.
What API Compatibility Lets You Reuse
Replay the same cases through Jev and Clef in shadow mode before changing routing. Keep identity checks and authorization in code. TypeSafe documents Jev 1.13 trouble with irrelevant long state, adversarial state and indirect criteria. Put those cases in the trial, and look for Clef-specific failures rather than assuming the lists match.
| Portable | Must be revalidated |
|---|---|
| Question names, answer definitions and state projection | Response-field equivalence in your client and version |
| Labeled cases and review rubric | Class errors, calibration and uncertainty meaning |
| Application action contract | Every allow, review and block threshold |
| Regression fixtures | Visual input behavior, provider failures and data handling |
How to Test Clef Against Jev
Cloudflare’s task list is useful for building a test set. Your held-out labels will tell you whether Jev, Clef or Clef-flash handles the queue better at the same review rate. Use the calibration guide for probability checks and Jev policies for the action boundary.
- Select one bounded task and label representative positive, negative and ambiguous cases without seeing model outputs.
- Use equivalent evidence and answer definitions through each documented native provider route.
- Keep raw distributions and all failures; fit thresholds separately on calibration data.
- Compare errors at equal review coverage, Brier score where probability semantics support it, latency and total workflow cost.
- Replay boundary cases, then shadow live traffic before changing routing or policy actions.
FAQ
Is Clef the same model as Jev?
No. Cloudflare trained Clef and Clef-flash and says the interface is Jev API compatible. The models and their measured behavior are separate.
Can I self-host Clef?
Cloudflare says it published Clef weights under Apache 2.0 on Hugging Face. Check the current repository, license and runtime requirements before deployment.
Is Clef more accurate or faster than Jev?
Cloudflare’s task table is mixed: Clef leads several rows, while Jev leads When2Call and BRIGHT. For your queue, run both through the intended provider path with the same labels and latency budget.
Sources
Checked against the sources below on October 2, 2026. Model versions, prices and limits change.
- Cloudflare: Introducing Clef
- Cloudflare Workers AI: Clef model and request examples
- TypeSafe AI: Jev primitives
- TypeSafe AI: Jev 1.13 jaggedness
Change note: Added Cloudflare’s Clef launch, Workers AI contract and benchmark caveats.