Cloudflare releases Clef-omni with audio and video support, cuts Clef-flash pricing

What happened

Cloudflare released Clef-omni, a multimodal decision model that handles audio, video, images, and text input in a single API call. Concurrently, Cloudflare reduced Clef-flash pricing to $0.038 per million input tokens—cheaper than TypeSafe's Jev—and optimized the base Clef model to run faster on its Workers AI platform.

Clef-omni's multimodal capabilities

Clef-omni natively processes audio in wav or mp3 format, video in mp4 or webm format, images, and text without requiring separate transcription or preprocessing pipelines. The model is built on Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts foundation already capable of multimodal processing. Cloudflare adapted the comprehension backbone while removing text-to-speech components, since Clef models are decision models rather than Large Language Models.

In production, Clef-omni performs a prefill pass across all input modalities simultaneously, then scores candidate values using two-stage attention routing. Video and audio are synced with visual frames for joint processing. Because the model skips output token generation, it avoids the overhead of transcribing or captioning incoming media. Clef-omni applies built-in lexical grammar to preserve option semantics and deliver schema-constrained scoring. Response latencies range from about 130 ms for text-only decisions to approximately 1.5 seconds for a full 21-second video clip with audio.

The model is trained using the same techniques as the original Clef: frozen Qwen3 backbone, low-rank adapters (LoRA), and post-training combining label-smoothed cross-entropy loss with Brier score calibration.

curl https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/cloudflare/clef-omni -X POST -H "Authorization: Bearer $CLOUDFLARE_AUTH_TOKEN" -H "Content-Type: application/json" -d '{"model": "clef-omni", "images": ["data:image/png;base64,<base64-png>"], "audio": ["data:audio/mpeg;base64,<base64-mp3>"], "videos": ["data:video/mp4;base64,<base64-mp4>"]}'

Clef-flash pricing and context-window trade-off

Clef-flash pricing fell to $0.038 per million input tokens from the previous $0.09, making it cheaper than Jev while maintaining performance across Cloudflare's benchmarks. To achieve this pricing, Cloudflare reduced the hosted context window from 64k to 24k tokens. Internal usage data shows that only 0.24% of requests exceed 24k input tokens, supporting this decision.

Developers with larger context requirements are encouraged to use the base Clef model, which remains at $0.24 per million input tokens and retains a 64k context window. The Clef-flash model weights on HuggingFace remain unchanged with full 256k context support for self-hosted deployments.

Clef model speed improvements

Cloudflare optimized the hosted Clef model's serving infrastructure without releasing new model weights. Median latency now measures 1.7× to 2.0× faster depending on input size: roughly 800-token inputs now return in 152 ms median (previously 262 ms), 3,400-token inputs in 305 ms (previously 616 ms), and 16,000-token inputs in 1,635 ms (previously 2,721 ms).

The primary optimization was migrating to SGLang for model serving. Cloudflare contributed pull requests to SGLang to incorporate Clef support, which will ship in SGLang 0.5.22. Updated self-hosting commands are available on the Clef HuggingFace repository and in SGLang cookbooks.

Real-world usage and availability

Cloudflare teams have deployed Clef for classification and detection across broad domains: spam issue detection in the public GitHub repository, phishing moderation in the EmDash content management system, personally identifiable information scanning for data loss prevention, and malicious domain detection by the threat intelligence team.

All three Clef models—Clef, Clef-flash, and Clef-omni—are open-weight, published on HuggingFace, and Jev-API compatible. Developers can switch models by changing the model ID via Cloudflare's AI Gateway. Full source code and developer documentation are available.

Source

AI-assisted, verified against the source. This post was drafted with AI from the primary source linked above and checked automatically against it. For anything you depend on, read the original.