Cloudflare Launches Clef-omni: A New Multimodal Decision Model

Cloudflare Introduces an Open-Source Multimodal Decision Model

Cloudflare has released a multimodal decision model called Clef-omni, licensed under Apache-2.0 and built on Alibaba's Qwen3-Omni-30B-A3B. The model's defining feature is its ability to score and evaluate content across four modalities — text, images, audio, and video — in a single API call, without requiring separate specialized models for each.

For use cases like content moderation and Trust & Safety, this unified decision-making capability translates to architectural simplification. Developers no longer need to build separate processing pipelines for different modalities; instead, they can consolidate judgment logic into a single model, reducing system integration complexity.

30B-A3B Architecture and Pricing

As its name suggests, Clef-omni inherits the 30B-A3B architecture from Qwen3-Omni. The "30B" refers to approximately 30 billion total parameters, while "A3B" indicates a sparse Mixture of Experts (MoE) structure with roughly 3 billion active parameters per inference. This design allows the model to maintain strong capabilities while only activating a fraction of its parameters at runtime — striking a balance between cost and performance.

Cloudflare has introduced a competitive pricing structure:

• Clef-omni: $0.15 per million input tokens
• Clef-flash: A lightweight variant priced at $0.038 per million input tokens

Clef-flash is approximately 75% cheaper than the standard version, making it well-suited for high-throughput scenarios where latency and cost sensitivity outweigh the need for maximum accuracy. The two-tier pricing gives teams the flexibility to choose based on their specific business requirements.

---

Mixture of Experts (MoE) is an architectural paradigm that divides a model into multiple "expert" sub-networks. During inference, a lightweight "router" dynamically selects a small subset of experts to participate in computation based on the input, rather than activating all parameters. This mechanism allows a model to maintain a large total parameter count (and thus broad knowledge capacity) while keeping the actual compute per inference comparable to a much smaller dense model. In the 30B-A3B case, while the model "has" 30 billion parameters, each forward pass activates only about 3 billion — delivering compute costs and latency similar to a dense 3B model, while retaining the knowledge representation capacity closer to a 30B model. This is the fundamental reason MoE architectures have been widely adopted in recent large language models, including Mixtral, DeepSeek^Ad, and the Qwen3 series.

---

The Value of Multimodal Decision Models

Traditional content judgment typically relies on classifiers trained for individual modalities — one for text, another for images, and separate ones for audio and video. When evaluating content that combines multiple modalities simultaneously, this fragmented architecture increases both call overhead and the difficulty of producing consistent, unified verdicts.

Clef-omni consolidates scoring capabilities across all four modalities into a single call, directly addressing this pain point. Paired with the permissive Apache-2.0 license, organizations can either use it as a pay-as-you-go managed service through Cloudflare, or self-deploy and fine-tune the open-source weights without commercial restrictions. This gives organizations that want full control over their data and models significantly more flexibility.

---

Trust & Safety is a core function for internet platforms managing user-generated content (UGC), covering scenarios such as policy violation detection, harassment identification, misinformation filtering, and CSAM (Child Sexual Abuse Material) interception. As rich media content — short videos, voice messages, mixed image-text posts — has exploded in volume, the limitations of single-modality classifiers have become increasingly apparent. For example, determining whether a video is policy-violating may require analyzing both the visuals and the audio together; analyzing either modality alone can lead to incorrect conclusions. Multimodal joint decision models, by fusing multiple input signals within a shared semantic space, can theoretically capture cross-modal semantic relationships more accurately — reducing both false negatives and false positives. This is the key reason such models are attracting attention from content safety teams.

---

Summary

Clef-omni represents a step toward unified, low-cost multimodal content decision-making. Its main selling points are the sparse Qwen3-Omni-30B-A3B architecture, integrated scoring across text, images, audio, and video in a single call, and accessible pricing at two tiers — $0.15 and $0.038. For teams building content moderation or Trust & Safety systems, this open-source model is worth adding to your technical evaluation shortlist.