Scores audio, video, image and text against typed questions in one call, returning probabilities instead of generated text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Single-Pass Scoring Across Four Modalities, Without Transcription Steps
Benchmarks
Decision Index: Mid-Table Overall, Strong Calibration
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
Clef-omni or Clef: Do You Need Audio and Video?
Clef and Clef-omni are sibling Cloudflare decision models. They share the System One API, typed question schema and probability outputs. The difference is input and backbone: Clef-omni adds audio and video on a 30B-A3B MoE, while Clef is a dense 27B model for text and images. Cloudflare's tables show Clef-omni beating Clef on four of ten tasks and trailing it on all five workflow scores. The decision is modality coverage versus text-task accuracy.
Choose Clef-omni when decisions depend on audio or video, or when you want to remove transcription and frame-splitting stages. It also carries a lower input rate than Clef. Choose Clef when inputs are text and images, especially for workflow-style tasks such as invoice processing and security incidents, where Cloudflare's results and the Decision Index Full score (53.08 versus 49.32) favor it. Both are listed with a 64K hosted context window.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Video and Audio Moderation
Clef-omni can score an uploaded clip against policy questions in one request. Videos are sampled at 2 frames per second, and the soundtrack is heard with the frames. That suits trust-and-safety checks that depend on what is both seen and said, with a probability you can threshold. Clips are capped at 60 seconds, so longer media needs chunking, and borderline scores should go to human review.
Multimodal Ticket and Intake Routing
Support and intake flows increasingly mix text, screenshots and voice notes. Clef-omni can route them with choice and score questions in one call. Cloudflare reports 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS, both vendor-run intent tasks. Use it when attachments matter. For text-only tickets, Clef or Jev may be more accurate on workflow-style evals.
Agent Guardrails and Gating
Place Clef-omni in the agent request path to answer yes/no or categorical questions before an LLM acts: block, escalate or proceed. Cloudflare reports about 130 ms median latency for text, and training adds Brier-score calibration, so outputs work as confidence values. Calibration error was 0.040 in the Decision Index run. It cannot take the action itself.
Equipment and Field Inspection Checks
Cloudflare's own example reviews an installation using a photo, an audio recording and a video, asking whether a label is visible, whether the unit sounds smooth and whether the fan is running. That cross-modal pattern suits QA, warranty and maintenance workflows. Each question returns a probability, and video length and file size caps apply per request.