Clef Omni

Clef-omni, released October 9, 2026 — eight days after the original Clef — is the third model in Cloudflare's decision family and the first that takes sound and video. It is a 30-billion-parameter mixture-of-experts model with ~3B active parameters, post-trained from Qwen3-Omni-30B-A3B-Instruct: Cloudflare kept the comprehension backbone and removed the text-to-speech components, so the model scores modalities and parameter options in a single prefill pass without generating output tokens.

One API call accepts text, JSON, images, wav/mp3 audio, and mp4/webm video, with audio and video streams synchronized to visual frames — removing the transcription and captioning stages multimodal agent pipelines usually need. Cloudflare reports ~130 ms for text-only decisions, ~150 ms with images, a few hundred ms for audio, and ~1.5 s for a 21-second video with sound, on a 64K-token context.

The benchmark picture is mixed and worth reading honestly. Cloudflare's own table shows 98.2 BFCL case-exact accuracy, 92.7 API-Bank, 94.8 BANKING77 macro-F1, and 97.7 CLINC150+OOS macro-F1 — but also has Clef-omni behind the text-only Clef on six of ten workflow evals (e.g. When2Call 63.3 vs 72.37), 2.6 accuracy on POP909-CL, and 47.4 on GPQA Diamond against Jev's 78.3. Pricing is $0.15 per million input tokens on Workers AI; weights are Apache 2.0, with self-hosting tested on a single H200 at ~64 GB in bfloat16.

Provider-reported: BFCL case-exact 98.2, API-Bank 92.7, BANKING77 macro-F1 94.8, CLINC150+OOS 97.7; behind Clef on 6 of 10 workflow evals; GPQA Diamond 47.4 vs Jev 78.3. Independent: none.