Cloudflare has announced a major expansion of its open-weight decision model ecosystem, introducing a new multimodal flagship model alongside substantial performance enhancements and price reductions across its existing lineup. Just a week after the initial rollout of Clef and Clef-flash, the company is continuing its rapid development cycle by launching Clef-omni, a model capable of processing audio, video, image, and text inputs within a single architecture.
The rapid deployment highlights Cloudflare’s emphasis on continuous iteration and swift engineering execution. According to the development team, the original concept for Clef was conceived on a Friday evening, trained over the course of a weekend, and brought to market by the following Thursday. Despite the compressed timeline, the initial releases delivered high-quality, performant open-weight models to the developer community. Building upon that momentum, the introduction of Clef-omni and the subsequent optimizations to Clef-flash and the core Clef model aim to make advanced decision-making tools more versatile, affordable, and accessible for modern agentic workflows.
Clef-omni Takes Audio, Video, Image, and Text Input
The debut of Clef-omni marks a significant shift in the capabilities of decision models, which have historically been limited to text-only formats since the introduction of comparable models like TypeSafe’s Jev. While the standard Clef model introduced support for images and video frame arrays, Clef-omni extends native ingestion to include direct audio files in WAV or MP3 formats and video files in MP4 or WebM formats, alongside traditional text and images.
This capability eliminates the need for complex, cascading engineering pipelines that previously required developers to transcribe speech to text or artificially split audio and visual channels from video containers. Instead, applications can invoke a single model to evaluate complex scenarios across multiple modalities simultaneously. This architecture brings automated decision systems one step closer to interacting with the real world through the combined mediums of audio, visual, and textual communication.
Under the hood, Clef-omni is built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. While the underlying model was already equipped to handle diverse modalities within a single pipeline, Cloudflare modified the architecture by removing the text-to-speech output components, focusing entirely on comprehensive input analysis. During execution, the model performs a rapid prefill pass across the complete payload, scoring all modalities and valid parameter options at the same time.
Because Clef models function as decision-making tools rather than traditional Large Language Models that generate open-ended text tokens, the architecture avoids the overhead associated with transcribing or captioning incoming media. Audio and video elements map directly into a unified sequence where visual frames and sound are synchronized for joint processing. Clef-omni then evaluates candidate values against internal embeddings through a two-stage attention routing mechanism. This process allows potential options to gather relevant evidence from any part of the input—whether buried in text, visible in an image, or heard in an audio stream—before cross-attending across the full context to compute final confidence scores. A built-in lexical grammar preserves option semantics, ensuring rapid, schema-constrained scoring regardless of the input type.
The training methodology mirrors the approach used for the original Clef model, relying on a frozen Qwen3 backbone adapted via low-rank adapters, followed by a post-training phase that combines label-smoothed cross-entropy loss with Brier score calibration. This design produces a robust system capable of handling schema variations, field ordering differences, and diverse prompt structures with high reliability.
Benchmark evaluations demonstrate that Clef-omni maintains strong performance across a variety of established tests and specialized workflows, including tool retrieval, API task execution, and domain-specific classification tasks. For developers interested in deployment, Clef-omni is available through Cloudflare developer documentation and its weights have been published openly on Hugging Face.
Clef-flash Model Is Now Cheaper
In response to developer feedback calling for decision models economical enough for integration into high-volume agentic workflows, Cloudflare has restructured the pricing for Clef-flash. Through targeted technical optimizations, the company has lowered the cost of Clef-flash so that it is now more affordable than competing options like Jev, while preserving high performance and output quality.
Achieving this more accessible price point required a specific architectural 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 open weights available on Hugging Face remain unchanged and continue to support a full 256k context window for developers who choose to self-host the model. Internal usage data indicated that only a fraction of a percent of requests—approximately 0.24%—exceeded 24k input tokens, justifying the adjustment to lower the barrier to entry. For applications requiring larger context handling, the standard Clef model retains its 64k context window.
Clef Model Is Now Faster
Alongside the introduction of new modalities and price adjustments, the standard Clef model has received significant speed improvements on the hosted Workers AI platform. These performance gains were achieved primarily through infrastructure-level optimizations rather than modifications to the underlying model weights.
A key factor in this acceleration was the integration of SGLang for model serving. Collaborating with the SGLang development team, Cloudflare contributed pull requests to incorporate Clef support, making the optimizations available in newer releases of the serving framework. Developers who prefer to self-host Clef can access the updated SGLang launch configurations through the official Hugging Face repository and SGLang documentation collections.
How Are People Using Clef?
Internal adoption at Cloudflare highlights the practical utility of domain-agnostic classification and detection models. Previously, deploying zero-shot classifiers across diverse domains often required specialized machine learning teams, custom training data corpora, and extensive fine-tuning to achieve acceptable accuracy. The Clef model family allows engineering teams to parse complex input data instantly through straightforward model calls.
Internal teams at Cloudflare have integrated the models into several critical production workflows. The public GitHub documentation repository utilizes Clef to automatically detect and close spam issues. The EmDash content management system employs the technology to scan plugin libraries for potential phishing threats. Additionally, data loss prevention teams rely on the models to scan sensitive data for personally identifiable information such as government identification numbers, while threat intelligence operations use Clef to aid in identifying malicious domains.
The expanded Clef model family, featuring native multimodal processing, reduced pricing, and accelerated execution speeds, remains fully compatible with the Jev API and accessible via AI Gateway. Cloudflare continues to encourage developers to explore the updated models and provide feedback as the ecosystem evolves.
Leave a Reply