Cloudflare has pushed a new artificial intelligence model onto its edge network, and it is not built to chat. On October 9, 2026, the company rolled out Clef-omni, a 30-billion-parameter multimodal decision model, onto Cloudflare Workers AI, giving developers a single model that reads text, JSON, images, audio and video and answers structured questions about all of it in one pass. The launch was first detailed in Cloudflare’s own developer changelog and picked up within hours by outlets including Ground News and AIbase, which flagged it as one of the more unusual AI releases of the week precisely because it skips the chatbot playbook entirely.
Clef-omni is accessible through the model ID @cf/cloudflare/clef-omni inside Workers AI, and it arrives alongside speed and pricing updates to two sibling models, Clef and Clef-flash. Together, the three form what Cloudflare calls its Clef family of decision models, a product line aimed squarely at the unglamorous but expensive work of content moderation, routing, scoring and guardrails, rather than generating prose, images or video.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Cloudflare Announced on October 9
The headline change is straightforward: Clef-omni adds native audio and video understanding to a model family that previously handled only text, JSON and images. According to Cloudflare’s changelog post, the new model accepts audio in WAV or MP3 format and video in MP4 or WebM format, on top of the text, JSON and image inputs its predecessors already supported. Cloudflare frames the release as expanding the Clef decision model family to “full multimodality,” while simultaneously making Clef faster and Clef-flash cheaper for existing customers.
That bundling matters for anyone already running moderation or classification pipelines on Workers AI. A model upgrade that also cuts the price of a smaller sibling model is a signal that Cloudflare is trying to make its entire decision-model lineup more attractive at once, not just launch a flagship and leave the rest of the lineup untouched. For a cloud platform competing with AWS, Google Cloud and Microsoft Azure on AI infrastructure, pushing three models forward in a single changelog entry is a notably efficient way to generate momentum.
What Is a Decision Model, and Why Clef-omni Isn’t a Chatbot
Most of the AI headlines of 2026 have been about models that generate things: text, images, video, code. Clef-omni does the opposite. Per its documentation on Hugging Face, the model takes a “state” (the content you want evaluated) and a “schema” of typed questions, then returns a calibrated probability for every allowed option of every question, in a single forward pass. There is no free-form text generation and, according to the model card, no output parsing required, because the response is already structured.
Think of the difference this way: a conventional multimodal chatbot might describe a video clip in a paragraph of prose that a developer then has to parse for a decision. Clef-omni instead takes a question like “does this video contain policy-violating content” with a fixed set of possible answers, and returns a probability distribution across those answers directly. That is a deliberate trade-off. Cloudflare’s own materials describe the officially stated use cases as classification, routing, scoring and guardrails over multimodal inputs, not open-ended conversation or content creation.
This “decision model” category sits closer to a classifier than a generative assistant, but it is multimodal and instruction-tuned in a way traditional classifiers were not. It is also a category Cloudflare has been building out gradually: Clef-omni is the third member of the family, following the original Clef and the smaller, cheaper Clef-flash, which Cloudflare introduced on October 1, 2026 as open-weight decision models hosted on Workers AI.
Inside the 30B-A3B Mixture-of-Experts Architecture
Architecturally, Clef-omni is a 30-billion-parameter mixture-of-experts (MoE) model with roughly 3 billion active parameters per inference, a design pattern often written in shorthand as “30B-A3B.” The idea behind MoE is that the full 30 billion parameters are never all active at once; instead, a routing mechanism selects a smaller subset, in this case about 3 billion parameters, for any given input. That keeps inference costs and latency closer to what you would expect from a much smaller dense model while retaining the broader knowledge capacity of a larger one.
This is the same broad architectural family used by several other 2026-era multimodal systems, which is part of why Clef-omni’s lineage matters. Cloudflare’s listing on Hugging Face states that Clef-omni was post-trained starting from Qwen3-Omni-30B-A3B-Instruct, the omni-modal foundation model built by Alibaba’s Qwen team. Qwen3-Omni-30B-A3B-Instruct is itself a natively end-to-end multilingual model with the same 30B-total, roughly-3B-active parameter split, supporting text, image, audio and video inputs. Cloudflare’s contribution was to take that general-purpose omni-modal base and fine-tune it specifically for the decision-model interface: typed schemas in, calibrated probabilities out.
That lineage is worth dwelling on because it reflects a broader pattern in how 2026’s AI infrastructure companies are shipping products. Rather than pretraining a frontier omni-modal model from scratch, which costs hundreds of millions of dollars in compute, Cloudflare built on an existing open-weight foundation model and specialized it for a narrower, commercially useful job. It is a faster and cheaper path to a production feature, and it is one more data point in the ongoing debate about how much differentiation actually comes from the base model versus the fine-tuning and serving layer built on top of it.
Four Input Types, One Forward Pass: How Clef-omni Reads Audio, Video, Image and Text
The practical appeal of Clef-omni is that developers no longer need to stitch together separate models for separate content types. Before this release, a team that wanted to moderate a platform with text comments, uploaded images, voice notes and short video clips would typically need at least two or three different specialized models, plus custom glue code to combine their outputs into a single decision. Clef-omni is pitched as collapsing that into one model and one API call, and Cloudflare reports roughly 20 ms for text requests, under 100 ms for image or audio inputs, and about 300 ms for a 21-second video with sound.
The supported format list is specific rather than exhaustive: WAV and MP3 for audio, MP4 and WebM for video, alongside standard image formats and plain text or JSON. That is a sensible starting set, covering the most common consumer upload formats, but it also means Cloudflare has made some conscious limits. A platform that primarily deals with less common audio or video codecs would still need a conversion step before calling the model.
Conceptually, the schema-and-state approach can be illustrated in a simplified way. A developer defines a question set describing what they want evaluated, and the model returns a probability per option for each question against the submitted content:
// Illustrative example of the schema concept, not official sample code
{
"state": { "video": "clip.mp4", "caption": "user-submitted text" },
"schema": {
"contains_policy_violation": ["yes", "no", "uncertain"],
"content_category": ["gaming", "music", "news", "other"]
}
}
// Response: a probability for each option, for each question, in one passThat structure is what Cloudflare means when it says the model needs no output parsing: there is no prose response to scrape for a verdict, just probabilities mapped directly onto the options a developer already defined.
Pricing and Access on Cloudflare Workers AI
Clef-omni is available now through Cloudflare Workers AI under the model identifier @cf/cloudflare/clef-omni, reachable through the same System One API that Cloudflare’s other Clef and Clef-flash models use, and compatible with Cloudflare AI Gateway for teams that want centralized logging, caching and rate-limiting across their model calls. Cloudflare’s own pricing documentation lists Clef-omni’s cost in both token and “neuron” terms, Cloudflare’s internal Workers AI compute-pricing unit, reflecting the model’s input-token processing cost. Workers AI lists Clef-omni at $0.150 per million input tokens, with a 64,000-token context window. Because Clef-omni returns structured probabilities rather than generated text, its pricing model is also structurally different from a conventional chat model: there is effectively no separate charge for long free-form output, since the model does not produce any.
That pricing structure reinforces the model’s core pitch. A generative multimodal model that happens to be used for moderation still bills for whatever token-heavy explanation it generates alongside its verdict. A dedicated decision model like Clef-omni is priced and designed around short, structured outputs, which should make cost more predictable for high-volume use cases like comment moderation queues or upload screening pipelines processing millions of items a day.
How Images, Audio and Video Count Against Input Tokens
Because Clef-omni bills per input token, the cost of a request depends on how each media type is converted into tokens. For images, 3 tokens are added to mark where the image starts and ends, and each image is capped at 1,024 tokens. Audio runs at about 780 tokens per minute.
Video is the heaviest input. It uses up to about 15,400 tokens per minute at maximum resolution, plus about 780 tokens per minute if the video has sound. A 480p clip uses about 8,600 frame tokens per minute (144 tokens per second).
These token rates sit against the $0.150 per million input tokens price and the 64,000-token context window listed for Clef-omni. For teams planning moderation pipelines, the practical step is to downscale video resolution where a policy decision does not need full detail, since resolution is what moves the frame-token count, and to check the per-minute token rates against the context window before sending long clips in a single call.
The Wider Clef Family Gets Faster and Cheaper Too
Clef-omni was not the only update in Cloudflare’s October 9 changelog. The company also cut the price of Clef-flash, its smaller and faster decision model, from $0.090 to $0.038 per million input tokens, while changing its hosted context window from the previously advertised 64K tokens to 24,000 tokens, and said it made inference on the original Clef model up to 2Ã faster. Clef itself stays at $0.240 per million input tokens. Cloudflare did not reposition Clef-flash as obsolete; instead, the three models now form a tiered lineup; Clef-flash for the lowest latency and cost on simpler text-and-image decisions, Clef for a faster general-purpose option, and Clef-omni for teams that specifically need audio and video understanding in the same decision pipeline.
That tiering mirrors how Cloudflare has shaped its broader Workers AI catalog, and it’s consistent with how the company has approached edge infrastructure pricing more generally, for instance in how it structured Cloudflare WAF’s free-to-paid tiers for web application firewall customers, or how its analytics engine stacks up in the Cloudflare Basin comparison against AWS Athena and BigQuery. Giving customers a cheap entry point alongside a premium option is a pattern across Cloudflare’s product line, not something invented for this release.
Why Enterprises Want Guardrails, Not More Chatbots
It is worth stepping back to ask why a decision model is news at all in a year dominated by headline-grabbing generative releases. The answer is that moderation, routing and guardrail workloads are enormous in volume and mostly invisible to end users. Every platform that accepts user-generated content, whether that is comments, forum posts, marketplace listings, voice messages or short video uploads, needs some layer that decides what is safe to publish, how to route a support ticket, or whether an upload matches a policy category. That layer typically runs on every single piece of content, which makes it far more volume-sensitive to cost and latency than a chatbot that only answers when a user asks a question.
A multimodal decision model that handles audio and video natively is attractive to exactly the companies dealing with that volume problem: social platforms, marketplaces, gaming communities with voice chat, and any SaaS product that lets users upload video. Cloudflare sitting at the network edge, where a large share of that traffic already passes through its infrastructure for DNS, CDN or DDoS protection reasons, gives it a plausible distribution advantage over a model that a developer would otherwise have to call out to a separate, more distant API.
This also connects to a wider security and compliance trend. Enterprises evaluating AI guardrail tools increasingly expect them to sit alongside existing security stacks, the same category of decision covered in comparisons like DNSFilter versus Cisco Umbrella versus Cloudflare Gateway and Zscaler versus Netskope versus Cloudflare One. A decision model that plugs into Cloudflare’s existing Gateway and AI Gateway products is as much a security and compliance pitch as it is an AI capability pitch.
Clef-omni vs Qwen3-Omni-30B-A3B-Instruct vs GuardReasoner-Omni
Clef-omni does not exist in isolation. Its own base model, Qwen3-Omni-30B-A3B-Instruct, is openly available and multimodal in its own right, and a separate academic approach, GuardReasoner-Omni, tackles a similar guardrail problem through multi-step reasoning rather than Clef-omni’s single-forward-pass probability output. The table below lays out how the three compare on the dimensions that are publicly documented.
| Model | Architecture | Input modalities | Primary output | Core purpose |
|---|---|---|---|---|
| Cloudflare Clef-omni | 30B-A3B mixture-of-experts | Text, JSON, image, audio, video | Calibrated probability per schema option | Classification, routing, scoring, guardrails |
| Qwen3-Omni-30B-A3B-Instruct | 30B-A3B mixture-of-experts | Text, image, audio, video | Free-form multilingual generation | General-purpose omni-modal assistant |
| GuardReasoner-Omni (research) | Reasoning-based multimodal guardrail | Text, image, video, audio | Reasoned safety verdict | Multi-step guardrail reasoning |
The pattern across all three is notable: the underlying omni-modal capability (reading audio, video, image and text together) has become common enough that multiple teams, commercial and academic, are building distinct products on similar architectural foundations. What differentiates Clef-omni from its own base model is entirely in the fine-tuning and interface layer; what differentiates it from GuardReasoner-Omni is the design choice of a single fast forward pass versus a reasoning chain, a classic latency-versus-depth trade-off that shows up across the guardrail-model space.
The Clef Model Family at a Glance
Here is how Cloudflare’s three current decision models stack up against each other following the October 9 update, based on Cloudflare’s own changelog and model documentation.
| Model | Input types | October 9 update | Positioning |
|---|---|---|---|
| Clef-flash | Text, JSON, image | Price cut | Lowest cost, lowest latency decisions |
| Clef | Text, JSON, image | Inference speed increase | General-purpose decision model |
| Clef-omni | Text, JSON, image, audio, video | New model launch | Full multimodal decisions, including audio and video |
All three remain accessible through Workers AI using the same System One API shape, which Cloudflare designed so that switching between tiers, say, starting a project on Clef-flash and later upgrading to Clef-omni once video support becomes necessary, does not require rebuilding the integration from scratch.
Historical Context: From Content Moderation APIs to Omni-Decision Models
Automated content decisioning is not a new problem. Text-based moderation classifiers have existed for well over a decade, and image-hashing systems for detecting known harmful content predate most of today’s generative AI boom by years. What has changed is the breadth of content types a single model can now handle, and the sophistication of the judgment it can apply, moving from “does this match a known bad hash” to “does this novel video clip, combined with its caption, violate a nuanced policy category.”
The jump to multimodal guardrails tracks the broader 2025-2026 shift toward omni-modal foundation models generally. Once base models like Qwen3-Omni could natively process audio and video alongside text and images, it became a matter of engineering effort, rather than fundamental research, to fine-tune a version of that capability specifically for structured decision-making. Cloudflare’s Clef-omni and the academic GuardReasoner-Omni project both arrived in roughly the same window of 2026, which suggests the industry reached a capability threshold where multiple independent teams converged on similar problems at a similar time, a pattern common throughout AI’s history whenever a new base capability becomes broadly available.
Cloudflare’s own AI push has been building for several years, moving from CDN and security services toward a full edge-compute and inference platform. Clef-omni is best understood as a continuation of that strategy rather than a one-off experiment: it extends an existing product line (Clef and Clef-flash) rather than introducing a new one, and it is released under the Apache 2.0 open-weight license, consistent with Cloudflare’s broader pattern of shipping open-weight models on its own platform rather than keeping them fully proprietary.
Market Impact: Cloudflare’s Edge AI Push Against AWS, Google and Microsoft
For Cloudflare, Clef-omni is a competitive move in the broader cloud-infrastructure fight as much as it is an AI product launch. AWS, Google Cloud and Microsoft Azure all offer their own content moderation and classification services, typically as separate, modality-specific APIs. Cloudflare’s pitch with the Clef family is unification: one model, one API shape, covering text, image, audio and video decisions, running on infrastructure already positioned at the network edge close to where user traffic originates.
Whether that unification argument wins developer mindshare will depend on factors beyond this single release: actual latency in production, accuracy on real-world edge cases, and how the pricing compares once teams run it at genuine scale rather than in a demo. But the strategic logic is consistent with how Cloudflare has approached other infrastructure categories, competing on simplicity and edge placement rather than trying to out-feature larger cloud providers point by point, a dynamic also visible in the serverless analytics space covered in the Cloudflare Basin comparison and in Cloudflare’s SASE security push against Zscaler and Netskope.
The release also lands at a moment when AI model pricing and capability is moving fast across the board. Just in the days surrounding this launch, Anthropic and other major labs have been adjusting pricing and model tiers, from Claude Opus 5.5’s benchmark results against GPT-6 Astra to new entrants like Grok 4.7’s coding-focused release. Cloudflare’s decision to specialize rather than chase frontier generative benchmarks is a distinct strategic lane, one that avoids directly competing with the handful of labs spending hundreds of millions on frontier pretraining.
Open Weights, Apache 2.0, and What It Means for Developers
Cloudflare published Clef-omni’s weights on Hugging Face under the Apache 2.0 license, one of the more permissive open-source licenses available, allowing commercial use, modification and redistribution with minimal restriction. That is a meaningful detail for developers who want to self-host the model outside Workers AI, run it on their own GPU infrastructure, or fine-tune it further for a narrower vertical, such as gaming voice-chat moderation or marketplace listing screening, without being locked into Cloudflare’s hosted pricing.
It also fits into a broader pattern across the decision-model and guardrail space: because the underlying judgment is narrower and more structured than open-ended generation, there is less competitive incentive for a vendor to keep weights closed. Cloudflare likely gains more from developers building on Clef-omni inside its ecosystem, where Workers AI billing and AI Gateway integration apply, than it loses from the model being inspectable or self-hostable elsewhere. OpenRouter already lists Clef-omni among its available providers, giving developers outside Cloudflare’s own platform another route to the same model.
What Comes Next: Five Predictions for Decision Models in 2027
- More audio- and video-native guardrails ship across competing clouds. Once one major edge or cloud provider ships a production multimodal decision model, rivals typically follow within a few quarters rather than years, especially given how widely available omni-modal base models like Qwen3-Omni now are.
- Pricing pressure keeps pushing decision-model costs down. The same October 9 changelog that introduced Clef-omni also cut Clef-flash’s price, suggesting Cloudflare expects competitive and volume pressure to keep compressing per-token costs across its whole decision-model tier, not just the newest model.
- Gaming and livestreaming platforms become an early adopter segment. Voice chat moderation and short-video screening are exactly the high-volume, multimodal workloads this category targets, and both sectors have struggled with scaling moderation cost-effectively.
- Expect more fine-tunes built on Qwen3-Omni and similar open base models. Clef-omni is unlikely to be the last commercial decision model built atop an openly available omni-modal foundation model rather than a from-scratch pretraining run.
- Academic guardrail research, like GuardReasoner-Omni, will increasingly cross-pollinate with commercial deployments. As reasoning-based guardrail approaches mature in research settings, expect commercial vendors to borrow reasoning-chain techniques for harder edge cases even inside otherwise single-pass systems.
How Developers Can Get Started
Teams already using Cloudflare Workers AI can call Clef-omni directly through the @cf/cloudflare/clef-omni model identifier, using the same System One API pattern as Clef and Clef-flash, which keeps migration friction low for anyone already on the platform. For teams not yet on Cloudflare, the model’s availability through OpenRouter and its open weights on Hugging Face under Apache 2.0 give two additional paths to evaluate it without an immediate platform commitment. Given that this is a specialized decision model rather than a general chatbot, the most productive first step for most teams is to map an existing moderation, routing or scoring workflow onto the state-plus-schema format before committing to a production integration.
For organizations already running AI agents or automated pipelines that touch sensitive data, this launch is also a reminder to revisit how model credentials and gateway access are secured, a topic covered in Tech Insider’s walkthrough of securing AI agent credentials through an MCP gateway. Adding another always-on multimodal model to a production pipeline is also one more reason to keep a close eye on access controls around it.
Frequently Asked Questions
What is Cloudflare Clef-omni?
Clef-omni is a 30-billion-parameter multimodal decision model that Cloudflare released on Workers AI on October 9, 2026. It takes text, JSON, images, audio and video as input and returns calibrated probabilities for typed questions, rather than generating free-form text.
How is a decision model different from a chatbot?
A chatbot generates open-ended text responses that a developer typically has to parse to extract a decision. A decision model like Clef-omni instead takes a predefined schema of questions and answer options and returns a probability for each option directly, with no free-form generation and no parsing required.
What base model is Clef-omni built on?
According to Cloudflare’s Hugging Face listing, Clef-omni was post-trained starting from Qwen3-Omni-30B-A3B-Instruct, an omni-modal foundation model developed by the Qwen team, which shares the same 30B-total, roughly-3B-active mixture-of-experts architecture.
What audio and video formats does Clef-omni support?
Clef-omni accepts audio in WAV or MP3 format and video in MP4 or WebM format, alongside standard image formats and text or JSON input, according to Cloudflare’s October 9 changelog post.
How do I access Clef-omni?
Clef-omni is available through Cloudflare Workers AI using the model identifier @cf/cloudflare/clef-omni, through the same System One API used by Clef and Clef-flash, and it is also listed as available through OpenRouter. Its weights are published on Hugging Face under the Apache 2.0 license.
Did Cloudflare update its other decision models at the same time?
Yes. Alongside the Clef-omni launch, Cloudflare said it reduced pricing for Clef-flash and increased inference speed for the original Clef model, as part of the same October 9, 2026 update.
What is Clef-omni meant to be used for?
Cloudflare’s officially stated use cases are classification, routing, scoring and guardrails over multimodal inputs, such as content moderation, support ticket routing, and policy enforcement across text, image, audio and video uploads.
Is Clef-omni open source?
Clef-omni’s weights are published under the Apache 2.0 license on Hugging Face, which permits commercial use, modification and self-hosting, separate from using the hosted version on Cloudflare Workers AI.
