Jev vs Clef
Multimodal schema head on Qwen, Jev-compatible request body.
Jev 1.13.0
TypeSafe · hosted System One model
77.1Capability · unranked reference
Intelligence 63.6 · calibration 90.6.
Clef
Cloudflare · Multimodal schema head on Qwen, Jev-compatible request body
63.8Clef-Flash · Capability #15
Two checkpoints. Clef-Flash is Capability #15 at 63.8. Clef 27B is 75.3 and outside the cost cap.
Key differences
- Where it runs JevHosted API, text inClefOpen weights; Workers AI also serves Clef by input token
- Licence JevProprietary, hostedClefApache-2.0
- Ranked checkpoint JevJev 1.13.0 reference, Capability 77.1ClefClef-Flash, Capability 63.8 · #15
- Larger checkpoint Jev—ClefClef 27B, Capability 75.3, intelligence 59.1, calibration 91.6, outside the cost cap
- Inputs JevText · images (experimental, 1.14)ClefText, JSON, images, or video — the board measured the text path
- Output JevTyped probabilities, no generated textClefOne logit per allowed option, softmax per question, no generated text
JevBench v1.6.1, same method
Capability is the mean of intelligence and calibration, ranked only inside Benchmark Heaven’s cost and median-latency caps. Jev 1.13.0 stays on the board as the reference row and is not given a rank. A model can score above that reference and still be left off the numbered list when its estimated cost is too high. Full table ↗
Capability axes, higher is better
Cost Flash $0.059 · Clef 27B $0.249 per 1,000 decisions (estimated). Median latency Flash 0.4 s · Clef 27B 0.6 s. Jev’s reference row is $0.032 per 1,000 decisions (not published as a list price), median 0.2 s. Both dollar figures are the benchmark’s, except Jev’s, which Benchmark Heaven marks as not a published list price.
Which Clef?
Clef is Cloudflare’s decision model, post-trained from Qwen3.8-27B with a joint schema head. A state and a schema of typed questions go in; every allowed option gets a probability in one forward pass. The request body matches Jev and System One, including choice, score, and yes/no, and the model card allows images and video on the request.
JevBench v1.6.1 ranks the smaller one. Clef-Flash is a Qwen3.5-9B post-train, Capability #15 at 63.8 (intelligence 42.2, calibration 85.5, about $0.059 per 1,000 decisions, 0.4 s). The 27B Clef scores Capability 75.3, intelligence 59.1, and calibration 91.6 — calibration above Jev’s 90.6 — at about $0.249 per 1,000 decisions, so it is measured and left unranked. Jev’s reference Capability is 77.1.
Read this before you switch
- The ranked Flash score is well below Jev on intelligence (42.2 vs 63.6). Use it for short, cheap text decisions, and read the 27B row before you treat “Clef” as one number.
- The board measured both checkpoints on text. Image and video behaviour is on the model card, not in that rank.
- Cloudflare bills Clef on input tokens. The JevBench dollar figures are the benchmark’s estimates, not the Workers AI invoice.
When to use which
Choose Jev if
- You want one hosted text model at Capability 77.1 without choosing between a 9B and a 27B.
- Intelligence on the independent board matters more than a multimodal schema head.
- You already call Jev and the requests are text.
Choose Clef if
- The decision depends on a video frame or audio, which Jev does not take — Jev 1.14 reads images experimentally, but Clef also takes video.
- You want Apache-2.0 weights and a joint schema head you can fine-tune (Unsloth documents a Clef-style head for that).
- You are already on Workers AI and want the Jev-compatible body there. Pick Flash for the cheaper ranked path, or the 27B when calibration 91.6 is worth the cost cap.
FAQ
Is Clef better than Jev?
The ranked Clef-Flash is not: Capability 63.8 vs Jev’s 77.1. The 27B Clef is closer, at 75.3, with higher calibration (91.6 vs 90.6) and lower intelligence (59.1 vs 63.6). It is unranked because estimated cost is $0.249 per 1,000 decisions.
Does Clef speak the Jev API?
Cloudflare’s model card says the Clef API takes a System One POST body and returns probabilities per question. It is a Cloudflare model, not a TypeSafe alias.
What is Clef-Flash?
The smaller, faster checkpoint: Qwen3.5-9B with the same joint schema head, multimodal weights, measured on text. It is the row with a Capability rank (#15, 63.8).