Cloudflare released Clef-omni on Friday, an open-weight model that handles text, images, audio, and video through one request instead of four. It is a decision model, meaning it scores a fixed set of options instead of writing text. The company’s blog post, published October 9, also reports a faster hosted Clef and a steep price cut for Clef-flash, with a catch on context length.

The multimodal part is the story. Today, a team that wants to judge a recording of a machine, a photo of it, and a clip of it running has to send each file type to a different model, or transcribe the audio first, then reconcile the answers. Cloudflare’s own example shows the alternative: one request carrying a photo of an installed unit, an audio recording, and a video of its fan, with three yes-or-no questions attached. Is the serial label visible, does the unit sound smooth, is the fan turning. One model scores all three.

Under the hood, Clef-omni starts from Alibaba’s Qwen3-Omni-30B-A3B, a mixture-of-experts model, with the speech-output half removed. It makes one pass over the whole payload and never generates tokens, which is why Cloudflare says it can skip captioning or transcribing the files first. Cloudflare’s reported latencies: about 130 milliseconds at the median for text, about 150 for an image, and about 1.5 seconds for a 21-second video that includes its soundtrack.

The accuracy picture is mixed, and Cloudflare’s table shows it. Against Clef, the text-and-image sibling, Clef-omni wins on API-Bank (92.7 versus 91.93), BANKING77 (94.8 versus 94.20), and CLINC150+OOS (97.7 versus 97.43). It loses on Home appliances (69.3 versus 82.95), When2Call (63.3 versus 72.37), and PhishNChips (73.2 versus 79.60). On the TypeSafe invoice workflow, its exact-action score is 60.2 against Clef’s 64.7. These are Cloudflare’s own runs, and none of the listed benchmarks is labeled as an audio or video test, so the table does not show how well the new inputs actually work.

The price list is where the post gets specific. Clef-flash falls from $0.09 to $0.038 per million input tokens, a cut of about 58 percent. Clef stays at $0.24, and Clef-omni launches at $0.15. The post gives no output-token prices for any of the three, and it says audio and image inputs are converted to input tokens by a formula it leaves to the docs. Until a team knows that conversion, $0.15 for a model that takes video is not a number it can compare with anything.

Clef-flash’s cut has a cost. The hosted version’s context window shrinks from 64k tokens to 24k. Cloudflare says its usage data shows 0.24 percent of requests run past 24k, and tells long-input users to move up to Clef. The published weights were trained for 256k, so self-hosters keep the room. The post says Clef-flash is now cheaper than Jev, TypeSafe’s decision model, but does not say what Jev costs.

Clef itself is faster, and here the post offers measurements. On the hosted service, median latency for roughly 800 input tokens fell from 262 to 152 milliseconds, for roughly 3,400 tokens from 616 to 305, and for roughly 16,000 tokens from 2,721 to 1,635. That is 1.7 times, 2.0 times, and 1.7 times faster. The baseline is Cloudflare’s own earlier hosted version, and the gain comes from serving changes, mainly a move to the SGLang inference engine, with no new weights. Tail latency improved less on short inputs: the 95th percentile went from 438 to 351 milliseconds at 800 tokens.

Why a network company is shipping decision models is a fair question, and the post does not ask it. Its examples suggest an answer: Cloudflare’s teams use Clef to close spam issues in a public docs repository, flag phishing in plugin libraries for its content system, scan for government IDs in data-loss prevention, and spot malicious domains. These are high-volume classification jobs that sit beside traffic Cloudflare already carries. The post does not say where hosted inference runs or address data residency; the open weights are the only route it offers for keeping data in a team’s own environment.

Cloudflare has shown that one model can take audio and video. It has not published a score for what the model does with them, and the only rows where Clef-omni can be compared with its text-and-image sibling include losses of roughly 13.7 points on Home appliances and 9.1 on When2Call.

Cloudflare, on its own blog, October 9, 2026.