Jev AIModel profile · Cloudflare · updated October 3, 2026
Cloudflare's open-weight decision models, with the same request shape as Jev and a vision encoder Jev does not have.
Give Clef a support ticket, a JSON record or up to four screenshots, plus the questions you need answered. It returns a calibrated probability for every allowed answer instead of prose, so your code can branch, route and escalate without parsing. Try Clef and Clef Flash below, then call either one with your Jev AI API key.
Cloudflare · 27B · Qwen 3.8-27B
The larger model. Leads Jev on eight of the ten decision benchmarks Cloudflare published, with a 209 ms median latency in its tests.
clefCloudflare · 9B · Qwen 3.5-9B
The fast model. Cloudflare reports a 39 ms median latency, with task-dependent quality differences such as weaker out-of-scope detection.
clef-flashThe same editor runs every model on Jev AI. Pick an example or start your own case, attach up to four images, edit the questions and run. Switch to Clef Flash from the model menu.
From Cloudflare's launch article and model card.
It scores, it does not write. Clef keeps a frozen Qwen backbone and adds rank-256 low-rank adapters plus a joint schema head: a transformer component that routes evidence from the state to each question and scores all allowed options in one forward pass. No tokens are generated, so there is no output to parse and no schema the model can break.
It is trained for calibration. Cloudflare combined label-smoothed cross-entropy on valid schema outputs with a Brier loss, then added a reinforcement-learning stage it calls Reinforcement Learning for Calibrated Decisions, which gives partial credit to adjacent answers. The probabilities are meant to be used as probabilities, not just ranked.
It sees. Both models carry a vision encoder. Up to four PNG, JPEG or WebP images are placed before the state, so a question can be about a screenshot, a product photo or a scanned document as easily as about text.
It has room. The documented context window is 65,536 tokens, twice the 32K Cloudflare quotes for Jev. Long text is truncated to fit, so keep each request focused and split large documents.
noulGet the probability that a condition is true, such as whether a ticket is urgent or a screenshot shows an error.
choiceRoute to an allowed team or label, with a probability for each option you define.
scoreDefine ordered levels and read a probability-weighted score, starting at zero.
Five of the ten decision benchmarks Cloudflare published on launch day, plus its latency figures. Vendor-reported, not measured by Jev AI. Metrics differ by row and are not averaged.
| Task / metric | Jev | Clef | Clef Flash |
|---|---|---|---|
| BFCLCase exact | 95.75 | 98.47 | 98.76 |
| When2CallAccuracy | 80.97 | 72.37 | 65.58 |
| BANKING77Macro-F1 | 79.74 | 94.20 | 90.93 |
| CLINC150+OOSMacro-F1 | 89.27 | 97.43 | 66.77 |
| PhishNChipsAccuracy | 62.55 | 79.60 | 75.05 |
| Median latencyms, lower is better | 524.1 | 209.3 | 38.8 |
| p95 latencyms, lower is better | 536.0 | 238.6 | 122.4 |
Clef leads on eight of the ten published rows; Jev keeps When2Call and BRIGHT. Clef Flash drops to 66.77 on CLINC150+OOS, so test out-of-scope detection before replacing a classifier with it. See all ten rows and the workflow evaluations.
Create an API key and store it as JEV_AI_API_KEY on your server. The endpoint, request and response are the same as for Jev; only model changes. Check authenticated GET /api/v1/models for configured models before sending requests.
curl --fail-with-body https://jev-ai.pro/api/v1/systemone \
-H "Authorization: Bearer $JEV_AI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "clef",
"state": "Checkout has failed for every customer for the last hour.",
"questions": {
"urgent": {
"type": "noul",
"instructions": "Does this incident need immediate attention?"
},
"team": {
"type": "choice",
"instructions": "Which team should handle this incident?",
"criteria": {
"billing": "Payments, invoices and refunds",
"technical": "Outages, errors and configuration"
}
},
"severity": {
"type": "score",
"instructions": "How severe is the customer impact?",
"criteria": [
"Minor",
"Major",
"Critical"
]
}
}
}'The response contains model, answers and usage. Read answers.<id>.noul, answers.<id>.probabilities and answers.<id>.score. Saved judges work with either model, and the Decisions API accepts the same images array.
Either way, keep the same state, questions and labels, run both models on held-out examples and compare against human-reviewed answers before changing a production route.
Clef is a family of decision models released by Cloudflare on October 1, 2026. Instead of generating text, a Clef model reads a state (text, JSON or images) and a schema of typed questions, then returns a probability for every allowed answer of every question. Clef is 27B parameters on a Qwen 3.8-27B base; Clef Flash is 9B on Qwen 3.5-9B.
No. Clef uses the same state-and-questions request shape as TypeSafe's Jev, so existing Jev requests run on Clef without changes, but it is a separate model with its own weights, calibration and limits. Jev AI never swaps one for the other: a Clef request stays on the Clef model you selected.
Yes. Both Clef models include a vision encoder. Through Jev AI you can attach up to four PNG, JPEG or WebP images per request, each up to 4 MiB and 16 megapixels and 8 MiB in total, in the playground or in the images array of the API. Remote image URLs are not accepted; images must be embedded.
Jev-Omni is an independent open-weights classifier that answers one Choice question about text, an image, audio or video. Clef answers up to 64 typed questions of three kinds in one call and reads text, JSON and images, but not audio or video. Both are available in the same Jev AI playground.
Start with Clef when accuracy matters most; Cloudflare's own table shows it ahead of Clef Flash on most tasks, including out-of-scope detection. Use Clef Flash when latency dominates and your evaluation shows its quality is sufficient. Both share the same request format, so switching is a one-field change.
Clef uses a 6× model rate and Clef Flash uses 3×. Playground runs cost 6 or 3 credits respectively. Token billing deducts actual input tokens multiplied by the model rate. API calls use tokens first, then the model’s credit charge for the whole request. Failed calls return the reserved balance. Cloudflare's upstream prices, $0.24 and $0.09 per million input tokens, are not Jev AI plan prices.
Reviewed October 3, 2026. Jev AI is an independent playground and API integration, not Cloudflare's official product site.
Clef, Qwen and every other product named on this page belongs to its respective owner and is not affiliated with Jev AI. Figures are quoted from the sources above as they read on the dates given and can change.