Cloudflare's Clef-omni: What Its Multimodal Scoring Model Actually Offers

Cloudflare says its Clef-omni decision model scores text, images, audio, and video in one call. The report says this can remove transcription and captioning steps that multimodal pipelines usually need. ¹ That could reduce orchestration work for some evaluation pipelines, but the benchmark numbers come from Cloudflare's announcement and have not been independently verified. ¹ For companies, the practical question is whether Clef-omni makes reliable decisions on their inputs at an end-to-end cost that compares favorably with their current workflow.

What the Clef-omni multimodal decision model is

Clef-omni is an open-weight multimodal decision model post-trained from Qwen3-Omni-30B-A3B-Instruct. It has 30 billion total parameters and 3 billion active parameters, and is released under Apache-2.0. ¹ Its practical pitch is scoring several media types through one call rather than requiring separate transcription and captioning steps. ¹

That could help when a decision depends on mixed evidence, for example judging whether a submitted clip meets a requirement based on its audio and visuals. This is a plausible workflow pattern, not a validated result. The report describes capabilities and benchmark scores, but does not establish performance on a particular company's data or provide production calibration results. ¹ A single-call design may reduce orchestration; it does not by itself prove that a decision is dependable or cheaper.

How scoring differs from generation

A decision model returns typed answers with probabilities instead of generated prose. Clef supports true-or-false, multiple-choice, and score-type questions. ¹ The goal is a signal software can act on directly: a category choice or a score against a rubric, rather than text a person has to parse.

Candidate uses include checking whether an image meets a visual specification, classifying an audio clip, flagging whether a video contains a specified event, or scoring an item against a written rubric. These are plausible applications of the stated capabilities, not reported results. For each task, a team still has to define the permitted answers, the acceptance threshold, and what happens when a result is ambiguous or low-probability.

A probability is not automatically a calibrated confidence score. The report says training combined label-smoothed cross-entropy with Brier score calibration, but does not report measured calibration performance. ¹ Analysis: before scores trigger consequential actions, compare predicted probabilities with observed correctness on held-out company data. Decide whether low-probability outputs go to a person, a fallback model, or a retry.

What the input limits mean for real workloads

On Workers AI, the report lists support for up to four images, four audio clips of up to 300 seconds each, two video clips of up to 60 seconds each, and between one and 64 questions per request. ¹ It says video frames and audio are synchronized into one sequence for joint processing before answer options are scored. ¹

Those limits are a starting point for test design, not evidence that every in-limit request will meet a particular quality or latency target. Check whether typical and worst-case inputs fit the stated bounds, and test whether splitting a longer item changes the decision. If a production request mixes media types, test that exact mix rather than extrapolating from single-modality examples. The report gives some latency measurements, but does not establish latency for a company's workload, accuracy on its own media, stability across repeated calls, or behavior on inputs that need preprocessing. ¹ Those remain questions to test.

Reading the Clef-omni benchmark table

Cloudflare's announcement reports the following Decision Index results for Clef-omni. The scores have not been independently verified. ¹

Benchmark	Clef-omni result	Comparison given	What it doesn't establish
BANKING77	94.8 macro-F1 1	No specific comparison score stated 1	Performance on a company's own categories or data
CLINC150+OOS	97.7 macro-F1 1	Reported as the top score among four Clef and Jev models 1	Performance under a company's own thresholds and error costs
Amazon ESCI	57.8 1	Reported as the top score 1	Fit for a specific ranking or scoring workflow
GPQA Diamond	47.4 1	Jev reported at 78.3 1	Performance on domain-specific tasks

The report also says Clef-omni scored below Clef on all five reported TypeSafe workflow metrics. ¹ That is a counterweight to focusing only on stronger results: reported performance varies by evaluation, and the scores have not been independently reproduced. ¹

Analysis: treat these figures as a reason to investigate, not a procurement verdict. They do not tell a buyer how the model handles their media mix, scoring rubric, decision thresholds, latency needs, or the relative cost of false positives and false negatives. A benchmark helps a purchase decision only when its task and conditions resemble the buyer's own. We make a similar point about another vendor result in Sakana AI's 73% core-claim result is a benchmark, not a deployment decision.

How to validate a multimodal AI model on your own inputs

Build a held-out evaluation set that mirrors the actual mix of text, images, audio, and video the product will receive. Include normal and hard cases, kept separate from anything used to design prompts or schemas. This is a suggested method, not a result reported by Cloudflare.

For each task, fix the question type (true-or-false, choice, or score) and write the permitted answers and acceptance rules before running the model. ¹ Then measure more than aggregate accuracy:

• Per-class performance, plus false-positive and false-negative rates.
• Whether predicted probabilities track observed correctness.
• Invalid or unusable outputs, and how often the workflow needs a fallback.
• Sensitivity to media length and consistency across equivalent presentations of the same input.

Run the current pipeline side by side on identical inputs, questions, thresholds, and review policy. Log preprocessing, retries, and human review alongside model outputs. The goal is to determine whether Clef-omni works for your task, not to name a universally best multimodal model.

Clef-omni pricing versus Clef-flash or an existing pipeline

The report lists Clef-omni at $0.15 per million tokens and Clef-flash at $0.038 per million tokens. ¹ It does not specify whether the Clef-omni rate covers input tokens only or how image, audio, and video inputs are counted. ¹ That leaves the listed rates incomplete for estimating the AI model input token cost of a real mixed-media request.

Option	Listed rate	Modality fit	Open billing questions
Clef-omni	$0.15 / million tokens 1	Verify on your actual mixed-media inputs	Token accounting, input-only scope, preprocessing or review overhead
Clef-flash	$0.038 / million tokens 1	Verify on the same task and inputs	No full workflow cost comparison provided
Current workflow	Not reported	Measure directly on the same inputs	Include preprocessing, model calls, retries, orchestration, and review

Analysis: compare end-to-end cost, not the model rate alone. Consider model charges alongside transcription or captioning, media conversion, retries, orchestration, storage or transfer, and downstream human review. The report does not provide figures for those items, so use your own usage data and vendor billing terms. Retries in particular can quietly change the bill; see testing for retries that cost money and outputs that vary.

A lower listed rate is not automatically cheaper if quality gaps increase retries, routing, or review. A higher rate could still pay off if one scoring call replaces several costly steps in a current workflow. Those are hypotheses to test against measured usage, not conclusions settled by the published prices.

Deployment: what's documented, what isn't

Cloudflare runs Clef-omni on Workers AI and reports self-hosting tests on a single H200 GPU, using about 64 GB of GPU memory in bfloat16. ¹ That is a reference point for infrastructure experiments; it does not establish production latency, throughput under a specific workload, scaling behavior, or quality parity between hosted and self-hosted versions. ¹

Analysis: validate in stages: schema handling and media acceptance, then quality and probability calibration, then latency and load under representative traffic, and finally end-to-end cost. Keep hosted and self-hosted results separate; measurements from one setup do not establish results for the other.

What to do next

Pick the narrowest workflow where single-call multimodal scoring might replace several evaluation steps. Before testing, define the decision, permitted outputs, quality threshold, and cost ceiling. Use the reported input limits to size test requests around the short and long media your team plans to send. ¹ When you shortlist alternatives to compare against, a model catalog is a reasonable place to start, as long as every candidate runs against the same evaluation set.

Ask Cloudflare about token accounting for non-text inputs, whether the Clef-omni rate is input-only, and how mixed-media requests are billed; the report does not specify those details. ¹ Treat the published benchmark scores as unverified until they are reproduced independently. ¹

The buying decision should rest on measured error behavior and end-to-end economics for your workload. Clef-omni's design and listed price are reasons to run a trial, not evidence that it is production-ready.

Sources

¹ “Meet Cloudflare Clef-omni: A 30B-A3B Multimodal Decision Model for Audio, Video, Image and Text.” MarkTechPost, October 11, 2026.