Clef-Flash

• Announcement: Clef decision models on the Cloudflare blog
• Decision Index leaderboard: clef-evals.workers-ai-mle.workers.dev

Clef-Flash is a 9B multimodal model that turns a state and a schema of typed questions into decisions. It reads the state as text, JSON, images, 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-Flash API is fully compatible with Jev and SystemOne.

Clef-Flash is post-trained from Qwen/Qwen3.5-9B. See Clef for the larger variant.

Model

• Backbone: Qwen/Qwen3.5-9B with its vision encoder, stored as standard sharded safetensors.
• 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.

Files

Usage

Tested with torch 2.11 and transformers 5.10.2 on a single H200. Image and video inputs also need pillow.

import sys

import torch
from huggingface_hub import snapshot_download

path = snapshot_download("Cloudflare/clef-flash")
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)))

SGLang

Clef-Flash can be served with SGLang. See the SGLang cookbook for launch commands on H200, B200, and B300. On a single H200:

docker run --gpus all \
  --shm-size 32g \
  -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --ipc=host \
  lmsysorg/sglang:dev-clef \
  sglang serve \
    --model-path Cloudflare/clef-flash \
    --host 0.0.0.0 \
    --port 30000

The server answers Jev/SystemOne requests at /v1/systemone:

curl http://localhost:30000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{"model":"Cloudflare/clef-flash","state":"My Stripe integration keeps failing. Please help ASAP.","questions":{"urgency":{"type":"noul","instructions":"Does this message express urgency?"}}}'

Jev / SystemOne API

systemone takes a Jev/SystemOne POST /v1/systemone request body and returns the same response body: model, answers keyed by question ID, and usage. A choice answer has choice, confidence, and probabilities; a score answer has the expected score, confidence, legend, and probabilities; a noul answer has the probability of true. instructions is optional, and images and videos may be added to the request.

from joint_schema_model import systemone

response = systemone(model, processor, {
    "model": "clef-flash",
    "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"])

Images and video

Add images (PIL images) or videos (frame arrays) to the record and pass the processor to encode_record. Optional processor arguments go in media_kwargs.

from PIL import Image

record = {
    "state": {"task": "Review the attached receipt."},
    "images": [Image.open("receipt.jpg")],
    "questions": {
        "legible": {"type": "noul", "instructions": "Is the receipt total legible?"},
    },
}
encoded = encode_record(processor.tokenizer, record, processor=processor)

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

Input format

Each question has:

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

encode_record accepts max_length (default 16,384 tokens) and max_state_tokens to bound the input.

Results

Decision Index

Per-benchmark results from our internal run of the Decision Index 0.2.1 suite. Scores are percentages; ForecastBench is a Brier score, where lower is better. The last two rows are request latency in milliseconds, where lower is better. The best value in each row is in bold.

Workflow evals

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

License

Released under the Apache-2.0 license, following the base model Qwen/Qwen3.5-9B.