Cloudflare Adds Clef-omni, Speeds Up Clef and Cuts Clef-flash Price

Clef-omni handles four input types in one call

The release follows last week's launch of Clef and Clef-flash, Cloudflare's open-weight decision models. According to Cloudflare, Clef-omni takes audio (wav or mp3) and video (mp4 or webm) in addition to text and images. The earlier Clef supported images and video frame arrays.

Cloudflare says this removes the need for cascading pipelines that transcribe speech to text or split audio and image channels from video. Developers can call one model to make decisions across any modality. The company published a curl example in which a single request combines a photo, an audio recording and a video, with yes/no style questions such as whether a label is visible or whether a fan is running.

The model weights are available on HuggingFace, and developer docs cover the API.

How Clef-omni works and how it performs

Cloudflare built Clef-omni on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. It keeps the comprehension backbone and discards the text-to-speech output components. Because Clef models are not Large Language Models, the company says they skip output token generation, so there is no overhead from transcribing or captioning incoming files. Training follows the approach used for Clef: a frozen Qwen3 backbone, low-rank adapters (LoRA), and post-training that combines label-smoothed cross-entropy loss with Brier score calibration.

On latency, Cloudflare reports text-only decisions in about 130 ms at the median and image inputs in about 150 ms. Audio clips take a few hundred milliseconds, and a full 21-second video clip with sound is scored in about 1.5 seconds, all in a single API call.

The benchmark tables in the announcement show a mixed picture. Clef-omni scores 98.2 on BFCL case exact, 94.8 on BANKING77 macro-F1 and 97.7 on CLINC150+OOS macro-F1. On several others, including When2Call (63.3 against 80.97 for Jev) and BRIGHT (42.0 against 47.52 for Jev), it trails. On the TypeSafe evals, Clef-omni and Clef are compared with Clef-flash and Jev across invoice processing, customer service, security incidents and agent trace observability.

Clef-flash gets cheaper, with a smaller context window

Cloudflare says it responded to feedback asking for a decision model affordable enough for any workflow. Pricing per million input tokens is now as follows:

• Clef-flash: $0.038, down from $0.09
• Clef: $0.24, unchanged
• Clef-omni: $0.15, launched today

The trade-off is the context window. The hosted Clef-flash now supports 24k tokens instead of the previously advertised 64k. Cloudflare says the weights on Hugging Face are untouched and were trained to support a 256k context window for those who self-host. According to the company's usage data, only 0.24% of requests exceed 24k input tokens. Clef keeps its 64k window, and Cloudflare recommends it for larger context needs.

Clef is up to 2.0x faster on Workers AI

The hosted Clef model on Workers AI is also faster. Cloudflare says no new weights were released, since most changes sit in the serving infrastructure rather than the model. One change was moving to SGLang. The company worked with the SGLang team on a pull request (PR #42721) to add Clef, which is due in SGLang 0.5.22.

Cloudflare's reported median and p95 latencies, before and now:

• About 800 tokens: 262 / 438 ms to 152 / 351 ms, a 1.7x median speedup
• About 3,400 tokens: 616 / 777 ms to 305 / 531 ms, 2.0x
• About 16,000 tokens: 2,721 / 3,250 ms to 1,635 / 1,805 ms, 1.7x

Self-hosting users can find updated SGLang launch commands on the Hugging Face repository and in the SGLang cookbooks.

What it means for developers

For developers building agentic workflows, Clef-omni offers a single endpoint for decisions over mixed media, while the lower Clef-flash price targets high-volume use. Teams with long inputs should check the reduced 24k window on hosted Clef-flash. Cloudflare points to its developer docs for current pricing, including how image and audio inputs are converted to tokens.