Model explainer · independent

What is Clef?

Clef is Cloudflare's family of open-weight decision models, released on 1 October 2026 during Birthday Week. Like Jev, it takes a state plus typed questions — noul, choice, score — and returns constrained answers with probabilities instead of generated text. Cloudflare calls the family fully Jev-API compatible.

This page is about Clef on its own terms: the variants, the three ways to run it, and what the published numbers actually establish. For the head-to-head with Jev, including the benchmark dispute, see the comparison page.

The family

Cloudflare describes the models as post-trained from frozen Qwen backbones with a routing head and low-rank adapters, rather than trained from scratch. The name is a musical clef — the symbol that fixes which note each line of a staff means — and the CF initials are not an accident.

VariantSize and backboneInputsPositioning
Clef27B, Qwen3.8-27BText, JSON, up to 4 imagesThe more precise model; classification-shaped tasks
Clef-flash9B, Qwen3.5-9BText, JSON, imagesThe latency-critical path; weaker on out-of-scope routing
Clef-omni30B-A3B mixture of experts, Qwen3-OmniAdds audio and video (2 frames per second)Media evidence; no published audio or video benchmark

Media details differ between model cards and the hosted API schema. Verify the exact request format on the surface you use. No audio or video accuracy figures have been published for Clef-omni.

Three ways to run it

Cloudflare also launched an RL fine-tuning platform alongside the models — AI Gateway captures traffic into a dataset, Containers sandbox rollouts, and the Trainer updates weights for redeployment on Workers AI. It is not required to use the released models.

  • →Workers AI serverless: call the hosted model IDs with a Cloudflare account; no GPU to operate.
  • →Self-hosted weights: download under Apache 2.0 and serve them yourself. Launch coverage describes Clef-flash on a single H200 and Clef-omni needing roughly 64 GB of GPU memory in bfloat16.
  • →Community runners: Ollama and similar tools list clef and clef-flash, which is the quickest local trial if you already have a runner installed.

What Cloudflare published, and what it did not

The launch post includes a 43-benchmark harness that Cloudflare selected and ran itself, including the competitor columns. In that run, Clef leads classification-style tasks such as BANKING77 (94.20 macro-F1 against Jev's 79.74) while Jev leads judgment-style ones such as When2Call (80.97 against 72.37). Three of the four workflow-eval margins are between 0.3 and 2.9 points.

The latency ordering flips in other hands. Cloudflare measured Clef at 209.3ms and Clef-flash at 38.8ms median against Jev at 524.1ms; an independent race through OpenRouter found Jev faster on every task. Community discussion disputed both the benchmark framing and which Jev build was tested. As of this review, no controlled neutral comparison has been published.

  • →Treat every accuracy score in the launch table as vendor-measured.
  • →Treat the price comparison as the cleanest number: $0.24 versus $0.042 per 1M input tokens on the hosted path.
  • →Test option-ID handling and required instructions fields before assuming a drop-in swap.
  • →Re-baseline confidence thresholds; an independent race found the two families calibrate quite differently.

Frequently asked questions

Is Clef open source?

+

The weights are Apache 2.0 on Hugging Face, so commercial use, self-hosting, and inspection are permitted. The training data and full pipeline are not published, so "open" here covers the artifact rather than full reproducibility.

Is Clef compatible with the Jev API?

+

Cloudflare describes the family as fully Jev-API compatible, and the request shape matches: state plus noul, choice, and score questions with probabilities returned. Field notes add cautions — Workers AI requires a per-question instructions field, option IDs affect results, and calibration differs. Test before swapping.

Is Clef faster than Jev?

+

It depends on who measured it. Cloudflare reports 209.3ms for Clef and 38.8ms for Clef-flash against 524.1ms for Jev in its own harness. An independent OpenRouter race found Jev faster on every task, and a Japan-side hands-on reported transit-inclusive numbers between the two. No like-for-like controlled result exists.

What is Clef-omni?

+

The third variant, reported on 9 October 2026: a mixture-of-experts decision model built on Qwen3-Omni-30B-A3B that accepts audio and video in addition to text and images. Use it when the evidence itself is a recording or a clip — but note no audio or video benchmark has been published.

How much does Clef cost?

+

On Workers AI, Cloudflare lists $0.24 per 1M input tokens for Clef and $0.09 for Clef-flash, with output free — roughly 5.7× and 2.1× the published Jev input rate. Self-hosting replaces the token bill with GPU time, electricity, and operations.

Keep reading

Sources

Reviewed 2026-11-13. Vendor- and project-reported figures are labelled as such; none were independently reproduced by this site.

Model explainer · reviewed 2026-11-13 · first published 2026-11-13