ENPT

Cloudflare adds Clef-omni (audio and video), cuts Clef-flash to $0.038/M and speeds up hosted Clef

Last updated:

Source · Model card: Clef / Clef-flash

On 9 October 2026 (20:11 BRT) Cloudflare announced three changes to its Clef family on its blog.

Clef-omni

A new open-weight model (Hugging Face, Apache-2.0) on Qwen3-Omni-30B-A3B-Instruct, with the speech-output parts removed. One request can carry text, images, audio (wav, mp3) and video with sound (mp4, webm). Same training recipe as Clef: frozen backbone, LoRA, label-smoothed cross-entropy plus Brier loss. Workers AI: @cf/cloudflare/clef-omni, $0.15 per M input tokens, 64K context; audio counts as about 780 tokens per minute.

Clef-flash price and Clef speed

  • Clef-flash: $0.09 → $0.038/M input, below Jev ($0.042). The hosted window drops from 64K to 24,576 tokens; weights unchanged (256K when self-hosted).
  • Clef: same weights and price ($0.24/M), now served on SGLang (PR #42721, due in 0.5.22). Cloudflare reports 1.7–2.0× lower median latency, e.g. 262 → 152 ms at ~800 tokens.

Numbers to read carefully

All speed and quality figures are Cloudflare’s (UnverifiedClaim). On its own ten-benchmark table Clef-omni trails Clef on most rows, so the gain is the new input types, not accuracy. The omni card’s Decision Index numbers come from an internal run of the superseded 0.2.1 suite. Clef-omni is not on JevBench v1.6.1 or Decision Index 0.3. Scores on the Clef page are unchanged.