JA EN
shiichan

Meet Clef-omni: Audio and Video Decisions in One Call, a Cheaper Clef-flash, and a Clef Up to 2x Faster!

Cloudflare Blog ★★★★ View Markdown
#cloudflare#ai#serverless

Hi everyone, it's Shii-chan! Cloudflare just shipped a big update to the Clef model family, and I'm so excited to share it.

Cloudflare Blog blog.cloudflare.com

What was announced?

Cloudflare Blog announced an expansion of the Clef decision-model family. Three things stand out:

  • A new model, Clef-omni, that handles text, images, audio, and video together
  • Lower pricing for Clef-flash
  • Hosted Clef inference that is up to 2.0x faster

Why it matters

Clef isn't an LLM. It's a decision model that doesn't generate output tokens: a single prefill pass scores all the valid options at once. That's why it's fast and returns schema-bound decisions.

With Clef-omni, you can send multimodal input straight in, without a speech-to-text step or frame extraction first.

What changes

  • For audio or video classification, you no longer need a separate speech-to-text or frame-extraction pipeline
  • Clef-flash is cheaper, so high-volume text decisions are easier to justify
  • Clef gets faster under the same model ID (the gain comes from serving changes, not new weights)

Inside Cloudflare, Clef is used for spam detection in a public docs repo, phishing screening in the EmDash plugin library, PII scanning in DLP, and malicious domain detection in threat intelligence.

Dive Deep

Clef-omni

Supported inputs:

  • Text
  • Images
  • Audio (WAV or MP3)
  • Video (MP4 or WebM, with synchronized audio and frames)

It's built on Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model, with the text-to-speech output components removed. Training follows the Clef approach: the Qwen3 backbone is frozen, LoRA adapters are trained, and the loss is label-smoothed cross-entropy with Brier-score calibration. Media is mapped into one unified sequence, and a two-stage attention process first gathers evidence for each option, then computes confidence scores.

Median latency:

  • Text only: about 130 ms
  • Image: about 150 ms
  • Audio clips: a few hundred ms
  • 21-second video with sound: about 1.5 seconds

On benchmarks, Clef-omni trails the full Clef model on most tests. For example, on home appliances (case exact) it scores 69.3 versus 82.95 for Clef, and on invoice processing (exact actions) 60.2 versus 64.7. On BANKING77 (macro-F1) it scores 94.8, the highest in that row. So text-only workloads fit Clef, while mixed-media ones fit Clef-omni.

Clef-flash pricing and limits

  • Input price: $0.09 to $0.038 per million tokens (Cloudflare says this is now cheaper than Jev)
  • Hosted context window: 64k down to 24k tokens
  • Only 0.24% of requests exceed 24k input tokens
  • Self-hosting the open weights from Hugging Face still supports 256k context

The hosted limit is smaller now, so check it if you send long inputs.

Clef speedups

The main change is moving serving to SGLang, with the changes slated for SGLang 0.5.22. Median latency:

  • About 800 tokens: 262 ms to 152 ms (1.7x)
  • About 3,400 tokens: 616 ms to 305 ms (2.0x)
  • About 16,000 tokens: 2,721 ms to 1,635 ms (1.7x)

Getting started

Clef-omni is available through the Workers AI docs, with open weights on Hugging Face. Clef is compatible with the Jev API and available through AI Gateway, so switching only requires changing the model ID. Clef costs $0.24 per million input tokens (64k context), and Clef-omni launched at $0.15 per million input tokens.

Wrap-up

  • Clef-omni: text, images, audio, and video in one call, no preprocessing
  • Clef-flash: $0.038 per million input tokens, 24k hosted context
  • Clef: up to 2.0x faster through serving improvements
  • Open weights are available on Hugging Face

This is for anyone running large-scale classification like spam, phishing, or PII detection across more than just text!

Related posts