Cloudflare posts Clef and Clef-flash, Apache 2.0 decision models on Workers AI

On October 1, 2026 at 19:51 UTC, @Cloudflare posted Clef and Clef-flash. The post calls them open-source decision models on Workers AI for classification and agent workflows, and it links the Birthday Week write-up. Michelle Chen is the byline on that post. The page says the weights are Apache 2.0 on Hugging Face, and that the hosted API accepts a Jev-shaped request.

Clef is the 27B model, post-trained from Qwen3.8-27B. Clef-flash is the 9B model, post-trained from Qwen3.5-9B. Both cards say the checkpoint keeps the base model’s vision encoder, adds a joint schema head, and returns one logit per allowed option. A softmax over each question turns those logits into probabilities. The cards say there is no free-form text to parse.

What you send

The hosted ids are @cf/cloudflare/clef and @cf/cloudflare/clef-flash. The docs show a Workers binding, env.AI.run, and a REST call:

POST https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/cloudflare/clef

The JSON body sets model to clef or clef-flash, plus state and questions. Question types are noul, choice, and score, the same three names Jev uses. The clef input schema allows 1 to 64 questions. A choice map is described as 2 to 255 options. A score list is 2 to 10 levels, lowest first. instructions is required on that schema. The local encode_record helper on the model card says instructions can be omitted, and that the question id is used instead.

The docs print a context window of 65,536 tokens for both models. The blog rounds that to 64k and sets it next to “Jev’s 32k”. The Workers AI card for typesafe/jev, already on Access, prints 32,000 tokens. TypeSafe’s own models page, as recorded on Access, says a Jev request is 64k tokens in total, and that state plus the longest question is 32k. The local helper’s default max_length is 16,384. The hosted window on the docs is 65,536.

Images on the hosted schema are an optional list of at most four embedded PNG, JPEG, or WebP items. Remote URLs are rejected. Each image is capped at 4 MiB and 16 megapixels, the decoded set at 8 MiB, and the whole body at 13 MiB. The model card also accepts video as frame arrays in a local record. The hosted schema we opened has no videos field. The docs intro still says the model reads video.

The clef card lists $0.24 per million input tokens. The clef-flash card lists $0.09. Neither page prints an output price. Jev’s published price on this desk is $0.042 per million input tokens, with output free. The blog says Cloudflare does not read, store, or train on requests or responses, unless you use the fine-tuning product described later on the same page.

The table they printed

The blog says Clef leads the Jev Decision Index on a live demo. We did not open that demo. The table under the sentence is ten rows. On seven of them the higher cell is Clef or Clef-flash. Jev is higher on When2Call and on BRIGHT. DiffusionGemma Jev is higher on PhishNChips.

Benchmark	Clef	Clef-flash	Jev	DiffusionGemma Jev	Kev 9B	Laya
BFCL, case exact	98.47	98.76	95.75	96.52	94.51	38.13
ToolRet, nDCG@10	69.19	66.43	65.28	61.21	64.26	12.69
API-Bank, accuracy	91.93	93.11	88.19	83.66	56.30	11.41
Home appliances, case exact	82.95	97.73	52.27	42.05	25.00	0.00
When2Call, accuracy	72.37	65.58	80.97	75.44	49.62	11.94
BANKING77, macro-F1	94.20	90.93	79.74	74.28	84.83	14.29
CLINC150+OOS, macro-F1	97.43	66.77	89.27	83.49	79.03	3.19
BRIGHT, nDCG@10	45.91	39.26	47.52	42.94	38.53	19.90
Amazon ESCI, macro-F1	57.48	57.39	55.21	53.37	49.22	24.40
PhishNChips, accuracy	79.60	75.05	62.55	85.35	50.75	50.15

Home appliances is the wide gap on that grid: Clef-flash 97.73, Clef 82.95, Jev 52.27. CLINC150+OOS is the wide miss for the small model: Clef-flash 66.77, Jev 89.27, Clef 97.43. BANKING77 here is macro-F1. Laya’s author table and Charly Poly’s encoder bake-off use other Banking77 setups. This page does not treat the three as one comparison.

A second blog table is four workflows, scored against TypeSafe’s eval suite. Invoice processing, exact actions: Clef 64.7, Clef-flash 57.1, Jev 61.8. Customer service: 76.3, 77, 76.0. Security incidents: 62.9, 61.7, 61.7. Agent trace observability: 68.5, 69.8, 71.6. The blog says the Clef models beat Jev in 3 of 4 areas. The invoice row is one of those areas for Clef, and Clef-flash is behind Jev there. Agent trace is the row Jev leads. The model card adds an invoice “primary action” line the blog table omits: Clef 86.2, Clef-flash 73.3, Jev 83.1.

Latency is one median and one p95, in milliseconds, which the blog says was taken across 43 evals. Median: Clef 209.3, Clef-flash 38.8, Jev 524.1, DiffusionGemma Jev 84.4, Kev 9B 51.4, Laya 5.8. p95: 238.6, 122.4, 536.0, 211.2, 187.9, 222.5. The blog says the Clef models are faster than the other decision models in that set except Laya. That exception matches the median row. On p95, Clef-flash’s 122.4 ms is under Laya’s 222.5 ms. 524.1 divided by 38.8 is about 13.5. 524.1 divided by 209.3 is about 2.5. Those ratios are arithmetic on the printed cells. The blog does not print them.

Where the longer card still has Jev ahead

The model card prints the same lab as an internal run of Decision Index 0.2.1, with scores rounded to one decimal. It marks a best cell in bold. Jev has that mark on eleven rows. Five of them, as Clef, Clef-flash, then Jev: GPQA Diamond 48.0, 51.0, 78.3. BBH 73.7, 68.9, 92.9. MMLU-Pro 65.9, 65.3, 82.7. When2Call 72.4, 65.6, 81.0. BRIGHT 45.9, 39.3, 47.5. The other six are ANLI, POP909-CL, VAST, NLI4CT, HoVer, and New Yorker.

CLINC150+OOS on that card is 97.4, 66.8, 89.3. RAGTruth hallucination F1 is 79.4, 35.6, 76.5. ForecastBench is a Brier score, lower better: 13.9, 10.6, 17.4. Clef-flash has the low cell on ForecastBench, and the low cell on RAGTruth.

This is not the composite on the Decision Index page. That page’s kit test locks Jev at 51.67, and the September 28 board file lists Jev 1.13.0 at 57.91. The Clef card does not print a number on that scale. Drex 1.5 and Liquid d1 argue over that composite. They are a different chart from these per-benchmark cells.

How they say it was trained

The blog says the week Jev shipped, the same team posted a DiffusionGemma experiment that read token probabilities. Michelle Chen’s September 18 post is that demo, at kev.workers-ai-mle.workers.dev, on Workers AI. The October post says Clef keeps the idea of scoring schema choices from a backbone, and that the backbone is Qwen. It points at Matt Mastracci’s vLLM pull 57250, which is already written up on DiffusionGemma as Jev.

The training note says Qwen3.8-27B and Qwen3.5-9B stay frozen. A routing head is trained together with rank-256 adapters. The loss they name is label-smoothed cross-entropy on valid schema outputs, plus a Brier term. The data they name is internal and synthetic, with field order, prompts, and schemas permuted. At inference the model runs a prefill, then scores the legal choices in parallel, without generating the answer token by token.

The same section says they developed Reinforcement Learning for Calibrated Decisions, RLCD, as a second target. The mechanics they list are partial credit for a neighboring ordinal choice, a reward for a fully correct record, and a penalty that keeps the distribution from drifting. This desk already uses RLCD as TypeSafe’s name for its own training, on the System One explainer. The Cloudflare post does not say the two procedures are the same one.

One internal example is on the blog, from Cloudflare’s threat-intelligence workflow. A domain plus a browser render is classified in 2.2 seconds. The post’s example distribution is about 95% fashion, 85% ecommerce, and under 1% phishing. The same workflow on gpt-oss-120b took 4.7 seconds and returned two labels. That is one workflow on their side. It is not a row in the table above.

Fine-tuning is a service, not a button on the model card. The blog says a forward-deployed team will tune Clef with a customer first, and that a self-serve trainer comes after that. The pieces they name are AI Gateway for capturing traffic, Workers AI for rollouts, Containers as a scoring sandbox, a new trainer, and bring-your-own-model for redeploying the result. The interest form is cloudflare.com/resource/clef-rl-interest. We did not submit it, and we did not call @cf/cloudflare/clef.

On October 3 Sahibzada Allahyar’s chart prints Clef at 61.2 and Clef-flash at 57.1 on Decision Index 0.2.1, with Jev at the public-board 57.91. The caption says the Cloudflare scores are self-reported, and the axis starts at 56. The model card we read still does not print a composite on that scale. 61.2 is not a replacement for the per-benchmark cells above, and it is not the blog’s latency row.

llama.cpp release v0.6.0, published October 5, serves these weights on POST /v1/systemone. Pull 29831 merged October 3 with commands for ggml-org/Clef-GGUF and ggml-org/Clef-Flash-GGUF, and said images were still missing. Pull 29969 merged October 5 and adds images for the clef type. The server README at that tag says a Clef process serves only that endpoint, and that Clef reads every question in one prompt. A choice can have 255 options. Opened October 6, the Hugging Face note lists Clef at 27B, images, Apache 2.0, and leaves the speed cell blank. Clef-flash is not a row there. The server write-up is llamacpp-jev. We did not run it.