Cloudflare shipped Clef-omni on October 9 — a 30B mixture-of-experts model that makes structured decisions about audio, video, images, and text in a single API call. It is not a chatbot. It does not generate prose. It reads a payload, answers a schema of yes/no and scoring questions, and costs $0.15 per million input tokens. Most developers building agents have never heard of the category. They should have.
Decision Models: The Tool Agents Actually Need
The default move when an agent needs to make a decision — is this content safe? should this ticket route to tier 2? does this audio sound like a distressed customer? — is to call a full LLM. This is wrong. LLMs generate text token by token. They are expensive, slow, and architecturally overkill for a yes/no answer.
Decision models work differently. They run a single prefill pass over the input, score a fixed set of structured answer options in parallel, and return probabilities — no decoding loop. Clef-omni’s 30B MoE architecture activates only about 3 billion parameters per inference, keeping latency under 200ms for most inputs and pricing closer to what you’d expect from a much smaller model.
What Clef-omni Adds: Audio and Video
The original Clef (27B) handled text, JSON, and images. Clef-omni adds audio (WAV and MP3) and video (MP4 and WebM). That might sound incremental. It is not.
The typical pipeline for audio-based decisions used to look like this: transcribe speech to text, run a text decision, discard the transcript. Each step adds latency, cost, and a point of failure. Transcription also destroys acoustic context — tone, hesitation, background noise — that is often exactly what you need for decisions like “does this customer sound distressed?” or “is this support call within policy?”
Clef-omni takes the audio directly. Video frames and the audio track are synced into a single joint sequence. One call, one decision, no intermediate pipeline.
const response = await env.AI.run("@cf/cloudflare/clef-omni", {
state: "Review the installation: a photo, audio clip, and video of the fan.",
images: ["data:image/png;base64,..."],
audio: ["data:audio/mpeg;base64,..."],
videos: ["data:video/mp4;base64,..."],
questions: {
label_visible: { type: "noul", instructions: "Is the label visible?" },
sounds_normal: { type: "noul", instructions: "Does the unit sound normal?" },
fan_running: { type: "noul", instructions: "Is the fan running?" },
},
});
The noul type returns a 0-to-1 probability for yes/no. The choice type accepts a list of valid options. The score type returns a numeric rating. All three can be combined in one schema, evaluated in a single pass.
Pricing: Clef-omni Is Cheaper Than the Model It Replaces
Clef-omni at $0.15 per million input tokens is 37.5% cheaper than the original Clef at $0.24/M — while adding audio and video. The same release dropped Clef-flash from $0.09 to $0.038/M (at the cost of a 24k context window; Cloudflare’s data shows only 0.24% of requests exceed 24k).
| Model | Price | Context | Modalities |
|---|---|---|---|
| Clef-flash | $0.038/M | 24k | text, JSON, images |
| Clef-omni | $0.15/M | 64k | text, JSON, images, audio, video |
| Clef | $0.24/M | 64k | text, JSON, images |
If you are currently using Clef for text and image decisions, Clef-omni is the obvious upgrade: same context window, lower price, more modalities.
Open Weights, Apache 2.0
The weights are on Hugging Face under Apache 2.0. Self-hosting requires about 64 GB of GPU memory in bfloat16. For Mac inference, mlx-community/clef-omni-4bit is the community quantized build. For vLLM: vllm serve "Cloudflare/clef-omni".
The caveat: open weights, not open source. Cloudflare has not released its training data or pipeline. But for regulated industries needing air-gapped deployments, or teams fine-tuning for narrow verticals like claims processing or call-center triage, the weights are what matters.
Cloudflare Uses It Internally — On Real Problems
Cloudflare runs Clef on production workloads: their GitHub docs repo uses it to auto-close spam issues, their EmDash CMS uses it to flag phishing in plugin libraries, the data loss prevention team scans for PII like government IDs, and threat intelligence uses it for malicious domain detection. These are not demos.
Which Clef to Use
Pick Clef-flash if your inputs are under 24k tokens and audio/video are not in scope — it is the cheapest option by a wide margin. Pick Clef-omni if you need audio, video, or the full 64k context window. There is no scenario where the original Clef is the right new choice today.
Decision models do not get the coverage that foundation model releases do. But they are the layer where most agent costs accumulate — and where most teams are currently overpaying. The full Clef-omni announcement and Workers AI API docs have everything you need to get started.

