↓ Skip to main content
9 – 17 UHR +49 8031 3508270 LUITPOLDSTR. 9, 83022 ROSENHEIM
DE / EN
DECISION Cloudflare USA

Cloudflare Clef (Clef & Clef-flash)

Cloudflare Clef, Clef-omni and Clef-flash: open decision models (Apache 2.0) with probabilities instead of text, Jev-compatible, from USD 0.038 per 1M input tokens.

License Apache 2.0
GDPR Hosting Available
Context 65536 Tokens
Modality Text, JSON, Image, Video, Audio → Typed decisions (choice, score, noul) with probabilities

Versions

Overview of available model variants

ModelReleaseEUStrengthsWeaknessesStatus
Clef-omni (@cf/cloudflare/clef-omni, 30B-A3B)
9 October 2026
Mixture-of-experts with 30B parameters (3B active), post-trained on Qwen3-Omni-30B-A3B-Instruct; the base model's speech output components are not used State as text, JSON, images, audio (WAV, MP3) and video (MP4, WebM) in one request; a video's soundtrack is evaluated together with its frames Limits on Workers AI: 64,000 token context, up to 4 audio clips (8 MiB and 300 seconds each) and 2 videos (16 MiB and 60 seconds each), audio and video together at most 16 MiB USD 0.15 per 1 million input tokens on Workers AI; audio costs about 780 tokens per minute, video up to about 15,400 tokens per minute plus sound Per Cloudflare the highest scores in the Clef family on BANKING77 (94.8), CLINC150+OOS (97.7) and Amazon ESCI (57.8); MMLU-Pro 70.6 per the model card Apache 2.0 on Hugging Face, Jev API compatible; self-hosting with Transformers, around 64 GB of GPU memory in BF16 (model card)
Below Clef and Clef-flash on tool tasks, for example 69.3 versus 83.0 and 97.7 in the home appliance simulator (model card) Hallucination detection (RAGTruth) 42.0 versus 79.4 for Clef per the model card SGLang support for self-hosting only announced so far per the model card Comparison figures from the vendor, language coverage not stated
Current
Clef (@cf/cloudflare/clef, 27B)
1 October 2026
Cloudflare's first in-house trained model: 27B parameters, post-trained on a frozen Qwen3.8-27B with vision encoder Multimodal state: text, JSON, up to 4 images (PNG, JPEG, WebP, 4 MiB and 16 megapixels each) and video; 65,536 token context, 1 to 64 questions per request Per Cloudflare benchmarks: BFCL 98.47, ToolRet nDCG@10 69.19, API-Bank 91.93, BANKING77 94.20 macro-F1, CLINC150+OOS 97.43 macro-F1 (Jev: 95.75 / 65.28 / 88.19 / 79.74 / 89.27) Jev API compatible (question types noul, choice, score, same response format); non-autoregressive, one logit per option USD 0.24 per 1 million input tokens on Workers AI, callable via Workers binding, REST /ai/run or Cloudflare AI Gateway Hosted, per Cloudflare 1.7 to 2 times faster since the serving optimisations of 9 October 2026 (including the move to SGLang): median 152 ms at about 800 tokens, 305 ms at about 3,400 tokens Apache 2.0 on Hugging Face (Cloudflare/clef); self-hosting with Transformers or SGLang (Docker image, Jev-compatible endpoint /v1/systemone) and a systemone() function Trained with label-smoothed cross-entropy, Brier loss for calibration and RLCD as a secondary target (per the blog)
Median latency 209.3 ms (p95 238.6 ms) on an H200 per Cloudflare, roughly five times slower than Clef-flash Knowledge benchmarks below Jev: GPQA Diamond 48.0 versus 78.3, MMLU-Pro 65.9 versus 82.7 (Cloudflare table) Self-hosting needs around 54 GB of GPU memory in BF16; community GGUF quantisations lack the decision head, PyTorch required (independent test report) All comparison figures come from Cloudflare; the only independent evaluation so far is a single test with 42 decisions No documented region pinning for Workers AI inference; language coverage not stated in the model card
Current
Clef-flash (@cf/cloudflare/clef-flash, 9B)
1 October 2026
9B parameters based on Qwen3.5-9B with vision encoder; median latency 38.8 ms, p95 122.4 ms on an H200 (Cloudflare), roughly 13 times faster than Jev in the same measurement USD 0.038 per 1 million input tokens on Workers AI (since 9 October 2026, previously USD 0.09), per Cloudflare cheaper than Jev; same input types as Clef (64 questions, 4 images, video) Per Cloudflare on par with or ahead of Clef on tool benchmarks: BFCL 98.76, API-Bank 93.11, BANKING77 90.93 Independent single test (42 agent decisions): 66.7% versus 71.4% for Jev, 83% agreement with the API, runs on an RTX 3090 Jev API compatible, Apache 2.0, self-hosting with far lower memory needs than Clef; per Cloudflare the weights were trained for up to a 256K-token context
CLINC150+OOS only 66.77 macro-F1 (Clef 97.43, Jev 89.27): out-of-scope detection markedly weaker (Cloudflare table) GSM8K 67.3 versus 80.8 for Clef per the model card; triage tasks in the independent test show the largest gap to Jev (10 of 20 versus 12 of 20) Hosted context limited to 24,576 tokens (Clef: 65,536); per Cloudflare 0.24% of requests exceed this limit. When self-hosting, the model card defaults to 16,384 tokens of input length Comparison figures from the vendor, language coverage not stated
Current

Use Cases

Typical applications for this model

Ticket and support triage with confidence
Classification of web pages, domains and bots (Cloudflare examples)
Trust and safety review of submissions
Decision nodes and supervision in AI agents
Image and video classification with typed answers
Decisions on voice recordings and videos with sound (Clef-omni)
Routing in front of expensive reasoning models

Technical Details

API, features and capabilities

API & Availability
Availability Public on Cloudflare Workers AI (binding env.AI.run, REST /ai/run, AI Gateway); weights on Hugging Face for self-hosting (Transformers, SGLang, Docker)
Latency (TTFT) 209.3 ms (Clef) or 38.8 ms (Clef-flash) median on an H200 (Cloudflare measurement); hosted Clef per Cloudflare 152 ms median at about 800 tokens
Features & Capabilities
Structured Output Vision
Training & Knowledge
Knowledge Cutoff not documented (backbones Qwen3.8-27B, Qwen3-Omni-30B-A3B and Qwen3.5-9B)
Fine-Tuning Available (Cloudflare RL fine-tuning platform (guided by forward-deployed engineers, self-serve announced), Your own fine-tuning of the open weights)
Language Support
Best Quality English
Supported Not stated; the Qwen backbones are multilingual, Cloudflare publishes no language evaluation
Cloudflare's benchmarks are in English; evaluate on your own German data before deployment

Hosting & Compliance

GDPR-compliant hosting options and licensing

GDPR-Compliant Hosting Options
Self-hosted
Your own EU infrastructure or EU cloud
Recommended for EU residency: Apache 2.0 weights, Transformers or SGLang, Clef around 54 GB and Clef-omni around 64 GB of GPU memory in BF16
Cloudflare Workers AI
Cloudflare global network
No reading, storing or training on customer requests per Cloudflare; no documented region pinning of inference in the Data Localization Suite
License & Hosting
License Apache 2.0
Security Filters None (decision model without text generation)
Enterprise Support Yes
On-Premise

Benchmarks

Performance comparison with standardized tests

BFCL case exact (Clef, Cloudflare figure)
98.47
BANKING77 macro-F1 (Clef, Cloudflare figure)
94.2
CLINC150+OOS macro-F1 (Clef, Cloudflare figure)
97.43
API-Bank accuracy (Clef-flash, Cloudflare figure)
93.11
MMLU-Pro (Clef, Cloudflare figure)
65.9

innFactory AI Consulting from Rosenheim adds Clef because Cloudflare extends the category of System One models we explained in Jev by TypeSafe: the AI model that writes no text with an open, multimodal and hosted offering. This page is based on the Cloudflare blog announcements for Clef and Clef-omni, the Workers AI changelog posts of 1 October and 9 October 2026, the model pages for Clef, Clef-omni and Clef-flash, the Workers AI price list and the model cards on Hugging Face. This overview reflects the state as of October 2026.

What is Clef?

On 1 October 2026 Cloudflare released Clef and Clef-flash, the first models the company has trained itself; Clef-omni followed on 9 October 2026. All three are decision models: they take a state (a support case, a web page, an agent trace, an image, a voice recording) plus a list of typed questions and return probabilities over predefined answers for each question. They generate no text. The question types are the same as with Jev: noul (yes/no as a probability), choice (one option from a list with a probability per option) and score (a level on a rubric). Cloudflare describes the models as fully compatible with the Jev API.

ClefClef-omniClef-flash
Workers AI ID@cf/cloudflare/clef@cf/cloudflare/clef-omni@cf/cloudflare/clef-flash
BaseQwen3.8-27B (frozen) with vision encoderQwen3-Omni-30B-A3B-Instruct with vision and audio encodersQwen3.5-9B (frozen) with vision encoder
Parameters27 billion30 billion, 3 billion active (MoE)9 billion
Inputstext, JSON, images, videotext, JSON, images, audio, video with soundtext, JSON, images, video
Context (Workers AI)65,536 tokens64,000 tokens24,576 tokens
Price (Workers AI)USD 0.24 per 1M input tokensUSD 0.15 per 1M input tokensUSD 0.038 per 1M input tokens
Median latency (Cloudflare, H200)209.3 msnot in the comparison table38.8 ms
LicenceApache 2.0Apache 2.0Apache 2.0

Technically, a joint schema head sits on top of the Qwen backbone: a small transformer that routes the evidence in the state to the questions and emits exactly one logit per option. Per the blog it was trained with label-smoothed cross-entropy, a Brier loss for calibration and Reinforcement Learning for Calibrated Decisions (RLCD) as a secondary target. The backbone stays frozen while low-rank adapters (LoRA) are trained; per Cloudflare, Clef-omni was trained with the same techniques.

Clef-omni: audio and video with sound

Clef-omni reads every modality in one request: text, JSON, images, audio (WAV or MP3) and video (MP4 or WebM). For videos the model evaluates the soundtrack together with the frames, so a decision can rest on what is seen and heard. Previously such a case needed a chain of transcription, track splitting and a text decision model. On Workers AI Clef-omni accepts up to 4 audio clips (8 MiB and 300 seconds each) and 2 videos (16 MiB and 60 seconds each); audio and video together may not exceed 16 MiB decoded. Billing runs through input tokens: audio about 780 tokens per minute, video up to about 15,400 tokens per minute at maximum resolution plus about 780 tokens per minute for the sound. Cloudflare gives different latency figures in the blog and in the changelog; before using it in latency-critical paths, measure it yourself.

Inputs and limits on Workers AI

  • State: text, JSON objects or arrays, up to 4 images (PNG, JPEG, WebP, at most 4 MiB and 16 megapixels each) and video; Clef-omni additionally accepts audio. Request body for Clef and Clef-flash at most 13 MiB, decoded image data at most 8 MiB.
  • Context: Clef 65,536 tokens, Clef-omni 64,000 tokens, Clef-flash 24,576 tokens. Images cost 64 to 1,024 input tokens depending on resolution; media tokens count toward the context, and an overly long text state is truncated.
  • Questions: 1 to 64 per request, question IDs up to 100 characters.
  • Response: model identifier, answers keyed like the questions, and usage.
  • Call: Workers binding env.AI.run(), REST endpoint /ai/run or via the Cloudflare AI Gateway; examples for TypeScript, Python and cURL in the docs.
  • Billing: Workers AI bills in neurons (Clef 21,818, Clef-omni 13,636, Clef-flash 3,455 neurons per 1 million input tokens); 10,000 neurons per day are free, beyond that USD 0.011 per 1,000 neurons on the Workers Paid plan. Cloudflare does not charge for output tokens on Clef models.

On 9 October 2026 Cloudflare cut the price of Clef-flash from USD 0.09 to USD 0.038 per 1 million input tokens and in return reduced the hosted context from 64K to 24K tokens; per Cloudflare 0.24% of requests exceed this limit. Anyone checking longer states uses Clef. At the same time Cloudflare moved the hosted serving of Clef to SGLang.

Benchmarks: Cloudflare’s figures

Cloudflare publishes a table against Jev in the blog; per the changelog a Clef model leads on 7 of 10 decision benchmarks. These are vendor figures; the GPQA Diamond and MMLU-Pro values come from the model cards.

BenchmarkClefClef-omniClef-flashJev (per Cloudflare)
BFCL, case exact98.4798.298.7695.75
ToolRet, nDCG@1069.1966.666.4365.28
API-Bank, accuracy91.9392.793.1188.19
BANKING77, macro-F194.2094.890.9379.74
CLINC150+OOS, macro-F197.4397.766.7789.27
GPQA Diamond (knowledge)48.047.451.078.3
MMLU-Pro (knowledge)65.970.665.382.7

Two things stand out. First, Jev leads clearly on the knowledge benchmarks, the Clef models on the tool and intent tasks. Second, Clef-flash collapses on CLINC150 with out-of-scope detection (66.77 versus 97.43 for Clef). Anyone picking Clef-flash for its latency or price should test out-of-scope cases separately.

In the workflow evaluations published by TypeSafe (invoice processing, customer service, security incidents) the models are close per Cloudflare: 64.7 / 60.2 / 57.1 / 61.8 on invoices, 76.3 / 71.6 / 77.0 / 76.0 in customer service, 62.9 / 61.7 / 61.7 / 61.7 on incidents (Clef / Clef-omni / Clef-flash / Jev).

Latency (Cloudflare measurement on an H200): Clef 209.3 ms median and 238.6 ms p95, Clef-flash 38.8 ms and 122.4 ms, Jev 524.1 ms and 536.0 ms. In the same table Cloudflare lists Laya at 5.8 ms median but 222.5 ms p95. For the hosted version of Clef after the serving optimisations, including the move to SGLang, Cloudflare states a median of 152 ms at about 800 tokens, 305 ms at about 3,400 tokens and 1,635 ms at about 16,000 tokens, 1.7 to 2 times the previous speed.

Independent single test: a developer compared Clef-flash locally on an RTX 3090 against the Jev API on 42 decisions of an autonomous agent: Jev 71.4%, Clef-flash 66.7%, the same decision in 83% of cases; identical on computer-use actions and supervision, 10 instead of 12 of 20 hits on triage. The test is small and not peer-reviewed, but it shows the order of magnitude. A practical note from it: community GGUF quantisations lack the decision head, PyTorch is needed for full functionality; Clef (27B) needs around 54 GB of GPU memory in BF16.

Data protection, regions and self-hosting

For the hosted variant Cloudflare states that requests and responses are neither read, stored nor used for training, unless the customer uses the fine-tuning offering. The Workers AI data usage docs confirm that customer content is not used to train models or improve services. Workers AI runs on Cloudflare’s global network; the Data Localization Suite documentation does not list Workers AI as a product with region pinning. Anyone who needs a contractual commitment to processing in the EU should clarify that with Cloudflare or take the second route.

The second route is self-hosting: the weights of all three models are on Hugging Face under Apache 2.0 (Cloudflare/clef, Cloudflare/clef-omni, Cloudflare/clef-flash). The model cards describe operation with Transformers and a systemone() function for the Jev-compatible interface; for Clef and Clef-flash SGLang is added, via a Docker image with a Jev-compatible endpoint /v1/systemone (per Cloudflare, support is due to land in the regular SGLang 0.5.22 release). For Clef-omni SGLang is announced. Cloudflare tested on a single H200 in BF16; per its model card Clef-omni needs around 64 GB of GPU memory. That lets you run the models on your own infrastructure or with an EU cloud provider, with full data control.

RL fine-tuning platform

Together with Clef, Cloudflare announced a platform for reinforcement learning fine-tuning. The building blocks: AI Gateway captures production data, Workers AI generates rollouts, Cloudflare Containers serve as the RL sandbox, a new Trainer updates the weights, and the result is redeployed via Workers AI with a bring-your-own model. At launch it is a guided offering with Cloudflare’s forward-deployed engineers via a design-partner form; a self-serve version is announced, pricing is not published.

Positioning: Clef, Jev and Laya

Clef / Clef-omni / Clef-flashJevLaya
LicenceApache 2.0, open weightsproprietary APIApache 2.0, open weights
OperationWorkers AI or self-hostingAPI only (US)self-hosting only
Size27B / 30B-A3B / 9B (Qwen base)not published0.3 to 0.4B (encoder)
Inputstext, JSON, images, video; Clef-omni also audiotext, JSONtext, JSON
Context (hosted)65,536 / 64,000 / 24,576 tokens64,000 tokens combined512 to 1,024 tokens, windows up to 8,192
Hosted priceUSD 0.24 / 0.15 / 0.038 per 1M inputUSD 0.042 per 1M inputno hosted API
EU residencyvia self-hostingnot documentedvia self-hosting

Clef closes the gap between Jev (strong, but closed and US-hosted) and Laya (open, but small): open models with an LLM-sized backbone, image and video input, audio input with Clef-omni, and a hosted option with a price. Hosted Clef-flash is priced below Jev, Clef and Clef-omni above it; per Cloudflare the latency of Clef-flash is below that of Jev. None of the three replaces an LLM; for text, summaries and code, models such as Qwen, Mistral or GPT-OSS remain in charge.

Our recommendation

For companies in the DACH region Clef is interesting above all as a self-hosted decision model: open licence, Jev-compatible interface, image and video input, and with Clef-flash a variant that runs on a single GPU. For decisions on voice recordings or videos with sound, for example in quality control or on call recordings, Clef-omni is the fitting variant. Hosted, Clef-flash is the cheapest of the three at USD 0.038 per 1 million input tokens, but limited to a 24,576-token context. Check three points before adopting it: out-of-scope detection with Clef-flash, language quality on German data (Cloudflare publishes English benchmarks only) and, for hosted operation, the contractual situation regarding the processing region. We assess its use as a front stage to language models in CompanyGPT or as a router in the AI Gateway per project. For an assessment of whether a decision model fits your architecture, contact innFactory AI Consulting.

Related decision models

The decision model (System One) category also includes Microsoft-Decision-1.1, Gemini Decision 1, Laya, Jev, Julia-1, GLiNER2.5-Decide. OpenAI is moving in the same direction with the Decisions API based on GPT-6 Luna; details are on the OpenAI GPT page.

Cost estimation for this model

For up-to-date token pricing, model variants and EU availability, see our sister project ai-prices.eu. It helps you compare and estimate the operational cost of leading AI models for your specific use case.

Compare prices on ai-prices.eu

ai-prices.eu is a project by innFactory AI Consulting GmbH and provides transparent cost estimates for leading AI models.

Frequently Asked Questions

What is Cloudflare Clef?

Clef is a decision model released by Cloudflare on 1 October 2026, the first model Cloudflare trained itself. It takes a state (text, JSON, images or video) and a list of typed questions and returns probabilities over predefined answers for each question, without generating text. Clef has 27 billion parameters based on Qwen3.8-27B, Clef-flash 9 billion based on Qwen3.5-9B. On 9 October 2026 Cloudflare added Clef-omni, a mixture-of-experts model (30B, 3B active) based on Qwen3-Omni that also processes audio. All three are available under Apache 2.0 on Hugging Face and hosted on Cloudflare Workers AI.

What is the difference between Clef, Clef-omni and Clef-flash?

Clef (27B) is the more accurate model, Clef-flash (9B) the faster and cheaper one, Clef-omni (30B, 3B active) the model for audio and video with sound. On an H200 Cloudflare measures a median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash. On Workers AI Clef costs USD 0.24, Clef-omni USD 0.15 and Clef-flash USD 0.038 per 1 million input tokens; in return, hosted Clef-flash is limited to a 24,576-token context, Clef to 65,536 and Clef-omni to 64,000. On CLINC150 with out-of-scope detection Clef-flash drops markedly to 66.77 versus 97.43 macro-F1 for Clef.

What does Clef cost on Workers AI?

The Workers AI price list states, per 1 million input tokens, USD 0.24 for Clef (21,818 neurons), USD 0.15 for Clef-omni (13,636 neurons) and USD 0.038 for Clef-flash (3,455 neurons); Cloudflare does not charge for output tokens on Clef models. Images, audio and video are converted into input tokens. Workers AI bills in neurons, 10,000 neurons per day are free, beyond that USD 0.011 per 1,000 neurons on the Workers Paid plan.

Can Clef be used in a GDPR-compliant way?

Two routes: hosted on Workers AI, Clef runs on Cloudflare's global network; Cloudflare states it does not read, store or train on requests and responses, but documents no region pinning for Workers AI inference in the Data Localization Suite. Anyone who must keep processing in the EU self-hosts the Apache 2.0 weights, for example with Transformers or SGLang on their own infrastructure or with an EU cloud provider; Clef needs around 54 GB of GPU memory in BF16, Clef-omni around 64 GB per its model card, Clef-flash considerably less.

How does Clef compare to Jev?

Per Cloudflare's own measurements Clef leads on 7 of 10 decision benchmarks, for example BANKING77 (94.20 versus 79.74 macro-F1) and BFCL (98.47 versus 95.75). On knowledge benchmarks Jev leads per the same table (GPQA Diamond 78.3 versus 48.0; MMLU-Pro 82.7 versus 65.9). An independent single-person test with 42 agent decisions saw Jev at 71.4 percent and Clef-flash at 66.7 percent with 83 percent agreement. Clef is compatible with the Jev API, so existing integrations can switch the base URL.

What is Cloudflare's RL fine-tuning platform?

Alongside Clef, Cloudflare announced a platform for reinforcement learning fine-tuning: AI Gateway captures data, Workers AI generates rollouts, Cloudflare Containers serve as the sandbox, a new Trainer updates the weights, and the result is redeployed via Workers AI with a bring-your-own model. At launch it is a guided offering with Cloudflare's forward-deployed engineers; a self-serve version is announced, pricing is not published.

Consultation for this model?

We help you select and integrate the right AI model for your use case.