Just a week after introducing its open-weight decision models, Cloudflare is expanding its artificial intelligence portfolio with the release of Clef-omni, alongside significant price reductions and performance enhancements across its existing lineup. The newly launched Clef-omni model extends the architecture’s capabilities beyond text and images, introducing native support for audio and video inputs. Alongside this multimodal expansion, Cloudflare has lowered the pricing for Clef-flash—making it more affordable than comparable industry offerings—and optimized its core infrastructure to deliver substantially faster processing speeds for the standard Clef model.
The rapid turnaround highlights an iterative development culture that has become central to Cloudflare’s engineering approach. The original concept for Clef materialized over the course of a single weekend, conceived on a Friday evening, trained over the next two days, and launched by the following Thursday. Despite the compressed timeline, the initial release delivered performant, high-quality, open-weight models to the developer community. Building on that momentum, the introduction of Clef-omni and the accompanying infrastructure improvements represent the next phase in the company’s broader push into the decision model space.
Clef-Omni Takes Audio, Video, Image, and Text Input
The debut of Clef-omni marks a notable shift for decision models, which have remained predominantly text-focused since the introduction of Jev from TypeSafe. While the original Clef release added support for images and video frame arrays, Clef-omni natively ingests audio files in WAV or MP3 formats alongside MP4 or WebM video, text, and images within a single request.
This capability eliminates the need for complex, cascading pipelines that stitch together separate speech-to-text transcription services or parse audio and visual channels independently from video files. By calling a single model, developers can process multiple modalities simultaneously to execute decisions. This unified approach brings the architecture closer to interacting with data across multiple sensory channels at once. Developers can access the new model through Cloudflare’s developer documentation or retrieve the open weights directly on Hugging Face.
Under the hood, Clef-omni is built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. While the underlying model natively handles text, imagery, audio, and video within a single pipeline, Cloudflare adapted the architecture by removing the text-to-speech output components. During execution, the model performs a rapid prefill pass across the complete payload, evaluating all input modalities and valid parameter options at the same time.
Because Clef models are designed for decision-making rather than acting as traditional large language models generating text outputs, the system avoids the overhead typically associated with transcribing or captioning incoming files. Media elements are mapped directly into a unified sequence, where audio and video streams sync with visual frames for joint processing. Clef-omni then extracts candidate values from internal embeddings using a two-stage attention routing mechanism, where every potential option gathers key evidence from text, visual, or audio streams before cross-attending across the full context to compute confidence scores. A built-in lexical grammar preserves option semantics, ensuring schema-constrained scoring across all input types.
Trained using frozen Qwen3 backbone weights combined with low-rank adapters and post-training techniques like label-smoothed cross-entropy loss paired with Brier score calibration, the model is built to maintain resilience against schema variations and prompt structures. In practice, the system handles text decisions in roughly 130 milliseconds at the median, image inputs in about 150 milliseconds, and audio clips within a few hundred milliseconds. Even a full 21-second video clip with accompanying sound can be scored in approximately 1.5 seconds through a single API call.
Extensive evaluations against established benchmarks demonstrate competitive performance across various workloads. In testing against TypeSafe evaluations, Clef models posted strong results across automated invoice processing, customer service automation, security incident evaluation, and agent trace observability workflows, maintaining high accuracy rates for exact and primary actions.
Clef-Flash Model Becomes More Affordable
In response to developer feedback calling for decision models cost-effective enough for widespread integration across automated workflows, Cloudflare has restructured its pricing for Clef-flash. Through architectural and operational optimizations, Clef-flash is now priced lower than competing models like Jev, incentivizing broader adoption across complex agentic pipelines.
Achieving this lower price point required a deliberate trade-off regarding the model’s context window. The hosted version of Clef-flash now features a 24k context window, down from the previously advertised 64k limit. However, the model weights available on Hugging Face remain unaltered and retain their original 256k context window capacity for users opting to self-host. Internal usage data indicated that only a fraction of requests—approximately 0.24 percent—exceed 24k input tokens, justifying the reduction to lower the barrier to entry. The standard Clef model retains its 64k context window for workloads requiring larger context handling.
Clef Model Achieves Faster Speeds
Performance optimizations have also resulted in speed improvements for the standard Clef model on Workers AI. Because these enhancements are concentrated at the serving infrastructure layer rather than structural modifications to the model weights themselves, no new weights were required for release.
Through these updates, median processing speeds have seen notable improvements across various input sizes. For inputs around 800 tokens, median processing times dropped from 262 milliseconds to 152 milliseconds, representing a 1.7-fold speedup. Payloads of approximately 3,400 tokens saw execution times fall from 616 milliseconds to 305 milliseconds, a twofold improvement. Larger inputs around 16,000 tokens dropped from 2,721 milliseconds down to 1,635 milliseconds.
A key factor in achieving these performance gains was the transition to SGLang for model serving. Cloudflare collaborated with the SGLang development team to integrate Clef support, with updates scheduled for release in SGLang version 0.5.22. Developers interested in self-hosting can access the updated launch commands and configuration guides through Cloudflare’s Hugging Face repository and official serving documentation.
Real-World Applications Across Internal Teams
Internal teams at Cloudflare have begun experimenting with Clef to handle domain-agnostic classification tasks. Previously, deploying zero-shot classifiers often required specialized machine learning engineering resources, domain-specific tuning, and extensive training datasets to achieve reliable performance. Clef enables developers to bypass custom training pipelines and immediately parse incoming data with high precision.
Internal deployments illustrate the versatility of the model family. The public GitHub documentation repository utilizes Clef to automatically detect and close spam issues. EmDash, Cloudflare’s content management system, employs the models to moderate plugin libraries for potential phishing threats. Additionally, data loss prevention teams scan corporate data streams for sensitive personally identifiable information, such as government identification numbers, while threat intelligence operations leverage the models to flag malicious domains.
As the Clef family continues to evolve with multimodal additions like audio and video integration, lowered cost barriers, and accelerated execution speeds, Cloudflare aims to make advanced decision models accessible through standard API integrations. Compatible with the Jev API and available via AI Gateway, the models require little more than a model identifier change for developers looking to incorporate automated decision-making into their existing applications.
Leave a Reply