One‑Model Multimodal Decision Making: What Clef‑omni Means for Your Daily Workflows

According to Cloudflare’s blog post, the company released Clef‑omni, a decision model that accepts text, images, audio, and video in a single API call. For anyone who currently stitches together separate transcription, OCR, or image‑classification services, this consolidation promises fewer moving parts and lower latency.

How Clef‑omni works under the hood

Clef‑omni builds on the Qwen3‑Omni‑30B‑A3B‑Instruct mixture‑of‑experts (MoE) foundation. An MoE model contains multiple expert sub‑networks and a router that decides which expert processes each token, allowing the system to handle different data types without blowing up compute cost. Cloudflare keeps the Qwen3 backbone for comprehension but drops the text‑to‑speech decoder, because Clef models are not meant to generate long passages— they only need to score a set of predefined answer options.

When you send a payload, the service runs a prefill pass that maps every modality into a unified token sequence. Video frames and audio samples are synchronized with visual tokens, then the model runs a two‑stage attention routing:

1. Evidence gathering – each candidate answer pulls key signals from any modality (e.g., a spoken keyword, an image label, or a piece of text).
2. Cross‑attention scoring – the gathered evidence is combined with a lexical grammar that enforces the schema of the decision task, and a confidence score is produced for each option.

Training follows the same recipe as the original Clef: the backbone is frozen, low‑rank adapters (LoRA) are added, and a loss that mixes label‑smoothed cross‑entropy with a Brier‑score calibration term is applied. The result is a model that tolerates variations in field ordering, missing modalities, or noisy prompts.

Speed and price at a glance

Clef‑omni’s claim to fame is that it does not require a cascade of services. The post provides raw latency numbers for each modality and a pricing table for the three Clef offerings. Below is a concise comparison:

The post also shows a speed boost for the hosted Clef model after infrastructure changes (SGLang serving): for a 3,400‑token request, median response time fell from 616 ms to 305 ms, a 2× improvement.

The hidden cost of a smaller context window

Clef‑flash is now advertised at $0.038 per M tokens, but the trade‑off is a reduced context window – the maximum number of input tokens the hosted service will accept. It dropped from 64 k to 24 k tokens. The model weights on Hugging Face still support a 256 k window for self‑hosting, so the limitation is purely a hosting‑layer decision.

Why does this matter? A 24 k token limit translates to roughly 15 k words, which is enough for most form‑filled requests but insufficient for long transcripts, full‑document reviews, or multi‑minute video analyses. If your workflow regularly pushes beyond that boundary, you’ll need to either:

• Switch to the full‑size Clef model (still $0.24 per M tokens) which retains a 64 k window, or
• Self‑host the flash weights and configure a larger window yourself.

Cloudflare’s usage data suggests only 0.24 % of requests exceed 24 k tokens, but the metric is skewed toward internal use cases. External teams handling compliance logs, call‑center recordings, or large PDFs should verify their typical token counts before committing to the cheaper tier.

Choosing the right Clef variant for your job

• If you need a single call that can look at a screenshot, an audio note, and a short text description – Clef‑omni gives you the simplest integration. Its latency stays under two seconds even for a 21‑second video, which is fast enough for real‑time ticket routing or chatbot assistance.
• If you only work with text or images and care about the lowest per‑token cost – Clef‑flash is attractive, provided your inputs stay under 24 k tokens. Remember the context reduction when planning batch jobs.
• If you regularly process long documents, multi‑minute recordings, or need the widest window without self‑hosting – stick with the original Clef model. It costs more per token but avoids the need to split inputs.

Try it today in under five minutes

1. Sign up for a Cloudflare AI Gateway account (free tier available).
2. Grab the API key from the dashboard.
3. Install the requests library if you haven’t already: pip install requests.
4. Use the following snippet, replacing YOUR_API_KEY and adjusting the files dictionary for the modalities you want to test:

import requests, json
url = "https://api.cloudflare.com/client/v4/accounts/<ACCOUNT_ID>/ai/run"
headers = {"Authorization": "Bearer YOUR_API_KEY"}
payload = {
    "model": "clef-omni",
    "prompt": "Select the correct category for the attached data.",
    "options": ["spam", "bug", "feature", "question"]
}
files = {
    "text": (None, "User reported error after playing the demo video."),
    "image": ("screenshot.png", open("screenshot.png", "rb"), "image/png"),
    "audio": ("note.mp3", open("note.mp3", "rb"), "audio/mpeg")
}
resp = requests.post(url, headers=headers, data=payload, files=files)
print(json.dumps(resp.json(), indent=2))

The call returns a JSON object with a confidence score for each option. Compare the latency you see with the numbers above; if it’s noticeably higher, check whether you’re hitting the 24 k token limit or whether the payload includes a full video.

By swapping the model field to clef or clef-flash, you can instantly benchmark cost versus speed for your own data. This hands‑on experiment will reveal whether the multimodal convenience outweighs the extra $0.11 per M tokens you’d pay for Clef‑omni.

---

Sources

• Cloudflare Blog – Introducing Clef‑omni with full multimodality, plus a faster Clef and a cheaper Clef‑flash