and… new clef drop from cloudflare!

https://t.co/gvb422RlpU

Cloudflare/clef-omni · Hugging Face

Source: https://huggingface.co/Cloudflare/clef-omni

• Announcement:Clef-omni model relase on the Cloudflare blog

Clef-Omni is a 30B-A3B mixture-of-experts multimodal model that turns a state and a schema of typed questions into decisions. It reads the state as text, JSON, images, audio, or video, and returns a probability for every allowed option of every question in a single forward pass. There is no free-form text generation and no output parsing.

The Clef-Omni API is fully compatible with Jev and SystemOne.

Clef-Omni is post-trained fromQwen/Qwen3-Omni-30B-A3B-Instruct. SeeClefandClef-Flashfor the dense image and video variants.

https://huggingface.co/Cloudflare/clef-omni#modelModel

• **Backbone:**the Qwen/Qwen3-Omni-30B-A3B-Instruct thinker with its vision and audio encoders, stored as standard sharded safetensors. The base model’s speech-output weights (talker and code2wav) are included unchanged but are not used;load\_release\_modeldoes not load them.
• **Joint schema head:**a small transformer head that reads the backbone’s final hidden states, routes evidence from the state to each question, and scores all options of all questions jointly.
• **Output:**one logit per allowed option for each question. Apply a softmax per question to get probabilities.

https://huggingface.co/Cloudflare/clef-omni#filesFiles

FilePurposemodel\-\*\.safetensors,model\.safetensors\.index\.json,config\.json,generation\_config\.jsonBackbone, including the vision and audio encodersjoint\_head\.safetensors,joint\_head\_config\.jsonJoint schema headjoint\_schema\_model\.pyRecord encoding, media decoding, batching, the model,load\_release\_model, andsystemone``tokenizer\.json,tokenizer\_config\.json,chat\_template\.jinja,processor\_config\.jsonTokenizer and image/audio/video processorLICENSEApache-2.0 license

https://huggingface.co/Cloudflare/clef-omni#usageUsage

Tested withtorch2.11 andtransformers5.10.2 on a single H200; the backbone needs about 64 GB of GPU memory in bfloat16. Image inputs also needpillow, and audio and video files needav.

import sys

import torch
from huggingface_hub import snapshot_download

path = snapshot_download("Cloudflare/clef-omni")
sys.path.insert(0, path)
from joint_schema_model import collate_records, encode_record, load_release_model

model, processor = load_release_model(path, device="cuda")

record = {
    "state": {"invoice": {"vendor": "Acme", "total": 1250.0, "currency": "USD", "status": "overdue"}},
    "questions": {
        "status": {
            "type": "choice",
            "instructions": "What is the invoice status?",
            "criteria": {"paid": "Invoice is paid.", "overdue": "Invoice is past due.", "draft": "Not sent."},
        },
        "large": {"type": "noul", "instructions": "Is the total above 1000 USD?"},
    },
}

encoded = encode_record(processor.tokenizer, record, processor=processor)
batch = collate_records([encoded], processor.tokenizer.pad_token_id, torch.device("cuda"))
with torch.inference_mode():
    logits = model(batch)[0]

for question, question_logits in zip(encoded.questions, logits):
    probabilities = question_logits.float().softmax(-1).tolist()
    print(question.question_id, dict(zip(question.option_ids, probabilities)))

https://huggingface.co/Cloudflare/clef-omni#jev–systemone-apiJev / SystemOne API

systemonetakes a Jev/SystemOnePOST /v1/systemonerequest body and returns the same response body:model,answerskeyed by question ID, andusage. Achoiceanswer haschoice,confidence, andprobabilities; ascoreanswer has the expectedscore,confidence,legend, andprobabilities; anoulanswer has the probability of true.instructionsis optional, andimages,audio, andvideos(base64 data URLs) may be added to the request.

from joint_schema_model import systemone

response = systemone(model, processor, {
    "model": "clef-omni",
    "state": "Our checkout started returning errors and orders are blocked.",
    "questions": {
        "department": {
            "type": "choice",
            "instructions": "Which team should handle the message?",
            "criteria": {"billing": "Payments or invoices", "technical": "Bugs or outages"},
        },
        "urgency": {"type": "score", "criteria": ["Can wait", "This week", "Today"]},
        "outage": {"type": "noul", "instructions": "Is a service down?"},
    },
})
print(response["answers"])

https://huggingface.co/Cloudflare/clef-omni#images-audio-and-videoImages, audio, and video

Addimages,audio, orvideosto the record and pass the processor toencode\_record. Each item can be a file path, URL, base64 data URL, or raw bytes. Images may also be PIL images, audio may be 16 kHz mono sample arrays, and videos may be RGB frame arrays sampled at 2 frames per second. Video files are sampled at 2 frames per second, and their soundtracks are heard alongside the frames when every video in the record has one. Optional processor arguments go inmedia\_kwargs.

record = {
    "state": {"task": "Review the attached call recording and dashcam clip."},
    "audio": ["call.wav"],
    "videos": ["dashcam.mp4"],
    "questions": {
        "glass": {"type": "noul", "instructions": "Is glass breaking in the audio clip?"},
        "collision": {"type": "noul", "instructions": "Does the video show a collision?"},
    },
}
encoded = encode_record(processor.tokenizer, record, processor=processor)

Text-only and multimodal records can be mixed in the same batch.

https://huggingface.co/Cloudflare/clef-omni#sglangSGLang

Use SGLang (coming soon!)

https://huggingface.co/Cloudflare/clef-omni#input-formatInput format

FieldDescriptionstateAny JSON value or string describing the situationimagesOptional list of imagesaudioOptional list of audio clipsvideosOptional list of videosmedia\_kwargsOptional keyword arguments for the processorquestionsMapping of question ID to question Each question has:

• type:noul(true/false),choice(named options), orscore(ordered options)
• instructions: what to decide; optional, and the question ID is used when it is omitted
• criteria: forchoice, a mapping of option ID to description; forscore, a list of option descriptions indexed from 0; fornoul, optional descriptions fortrueandfalse

encode\_recordacceptsmax\_length(default 64,000 tokens) andmax\_state\_tokensto bound the input.

https://huggingface.co/Cloudflare/clef-omni#resultsResults

https://huggingface.co/Cloudflare/clef-omni#decision-indexDecision Index

Per-benchmark results from our internal run of theDecision Index0.2.1 suite. Scores are percentages; ForecastBench is a Brier score, where lower is better. The best value in each row is in bold.

BenchmarkClef-OmniClefClef-flashJevBFCL (case exact accuracy)98.298.598.895.8ToolRet (nDCG@10)66.669.266.465.3API-Bank (accuracy)92.791.993.188.2BANKING77 (macro-F1)94.894.290.979.7CLINC150+OOS (macro-F1)97.797.466.889.3RouterBench (selected quality)79.979.779.9****79.9Home appliance simulator (case exact accuracy)69.383.097.752.3SGD/SGD-X (macro-F1)39.243.834.243.0ContractNLI (macro-F1)80.381.484.371.7ANLI (macro-F1)65.569.859.174.8BPoMP (accuracy)96.896.995.490.6Humicroedit (accuracy)70.166.775.161.9POP909-CL (accuracy)2.615.81.618.1cfcolor (accuracy)62.966.065.864.7MMLU (accuracy)92.790.391.891.7GPQA Diamond (accuracy)47.448.051.078.3ARC-Easy (accuracy)99.899.099.599.3ARC-Challenge (accuracy)97.897.798.397.8WinoGrande (accuracy)96.793.597.592.0HellaSwag (accuracy)99.198.298.694.5GSM8K (accuracy)79.080.867.379.9ChessBench (accuracy)22.124.723.017.2MuSR (accuracy)83.883.586.066.1SATA-Bench (case exact accuracy)31.933.836.726.4BRIGHT (nDCG@10)42.045.939.347.5Amazon ESCI (macro-F1)57.857.557.455.2ACOS (per-review F1)28.433.325.929.5FinEntity (macro-F1)96.396.297.187.0VAST (macro-F1)50.959.549.664.6NLI4CT (macro-F1)76.882.978.684.1CRUXEval (accuracy)88.886.786.173.0CLadder (accuracy)97.294.097.772.6ForecastBench (Brier, lower is better)11.713.910.617.4Habermas Machine (accuracy)67.868.771.845.9PhishNChips (accuracy)73.279.675.062.5MMLU-Pro (accuracy)70.665.965.382.7BBH (accuracy)69.073.768.992.9RAGTruth (hallucination F1)42.079.435.676.5HoVer (accuracy)64.265.261.272.9When2Call MCQ (accuracy)63.372.465.681.0New Yorker (accuracy)64.269.566.170.1

https://huggingface.co/Cloudflare/clef-omni#workflow-evalsWorkflow evals

Decision accuracy on four end-to-end business workflows fromTypesafe Evals, scored against consensus reference labels. All models are scored on the same dataset revision and case cohort.

WorkflowMetricClef-OmniClefClef-flashJevInvoice processingExact actions60.264.757.161.8Invoice processingPrimary action82.086.273.383.1Customer serviceExact actions71.676.377.076.0Security incidentsExact actions61.762.961.761.7Agent trace observabilityPrimary action65.868.569.871.6

https://huggingface.co/Cloudflare/clef-omni#licenseLicense

Released under the Apache-2.0 license, following the base modelQwen/Qwen3-Omni-30B-A3B-Instruct.