Just a week after introducing its initial open-weight decision models, Clef and Clef-flash, Cloudflare is expanding its artificial intelligence portfolio with the release of Clef-omni. The new model introduces native support for audio and video inputs alongside text and images, marking a significant evolution for decision architectures that have traditionally relied solely on text processing. Alongside the launch of Clef-omni, Cloudflare has lowered the pricing for Clef-flash—making it more affordable than comparable industry offerings like Jev—and implemented infrastructure-level speed enhancements that significantly accelerate the original Clef model.
The rapid development cycle underscores Cloudflare’s ongoing commitment to rapid iteration and deployment. The foundational concept for Clef materialized over the course of a single weekend, hatched on a Friday evening, trained across Saturday and Sunday, and launched by the following Thursday. Despite the compressed timeline, the engineering team delivered high-performance, open-weight models for the broader developer community. Building on that momentum, the introduction of Clef-omni and the accompanying optimizations aim to make decision models faster, more cost-effective, and more versatile for agentic workflows across the industry.
Clef-omni Takes Audio, Video, Image, and Text Input
The launch of Clef-omni fundamentally alters the operational paradigm for decision models, which have remained largely text-centric since the debut of TypeSafe’s Jev. While the original Clef model introduced support for images and video frame arrays, Clef-omni expands these capabilities to ingest audio files in WAV or MP3 formats and video files in MP4 or WebM formats directly alongside text and static images.
This multimodal ingestion eliminates the need for complex, cascading engineering pipelines that typically stitch together separate speech-to-text transcription models or split audio and visual channels from video streams. Instead, developers can call a single model to evaluate decisions across multiple modalities simultaneously. This brings automated systems a step closer to interacting with the world through the same combination of visual, auditory, and textual communication experienced by humans.
The new model is available via developer documentation on Cloudflare Workers AI and has been published with open weights on Hugging Face. Clef-omni is built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, which natively processes text, imagery, audio, and video within a unified architecture. Cloudflare adopted this core comprehension backbone while omitting the text-to-speech output components typically associated with large language models. During execution, Clef-omni performs a rapid prefill pass across the complete payload, evaluating all modalities and valid parameter options at the same time.
Because Clef models function as decision engines rather than generative large language models, they bypass output token generation entirely. This design removes the computational overhead normally required for transcribing incoming files or generating descriptive captions. Media elements map directly into a unified sequence where video frames and audio tracks are synchronized for joint processing. Clef-omni then extracts candidate values from internal embeddings using a two-stage attention routing mechanism. Every valid option gathers relevant evidence from the input data—whether located in text snippets, visual frames, or audio streams—before cross-attending across the full context to compute exact confidence scores. A built-in lexical grammar preserves option semantics, ensuring rapid and schema-constrained scoring across every input format.
The training methodology mirrors that of the original Clef model, utilizing a frozen Qwen3 backbone combined with low-rank adapters and a specialized post-training approach that pairs label-smoothed cross-entropy loss with Brier score calibration. This training results in a resilient architecture capable of maintaining accuracy despite schema variations, differing field orderings, and diverse prompt structures. Consequently, calibrated and schema-bound decisions can be executed across text, images, audio, and synchronized video through a single API call. Performance testing demonstrates rapid execution times, with text-only decisions returning at a median of approximately 130 milliseconds, image inputs processing in about 150 milliseconds, and audio clips taking a few hundred milliseconds. Even a comprehensive 21-second video clip with accompanying sound can be fully scored in roughly 1.5 seconds.
Clef-Flash Model Is Now Cheaper and More Accessible
In response to developer feedback calling for more economical decision models suitable for widespread deployment across agentic workflows, Cloudflare has recalibrated its pricing structure. Through targeted engineering optimizations, Clef-flash is now priced lower than Jev while preserving its performance and quality standards.
To achieve this reduced price point, Cloudflare made a deliberate trade-off by adjusting the hosted context window for Clef-flash. The hosted version now operates with a context window of 24,000 tokens, down from the previously advertised 64,000 tokens. However, the open weights hosted on Hugging Face remain unaltered and continue to support the full 256k context window for developers choosing to self-host the model. Internal usage data indicated that only a fraction of a percent of requests—specifically 0.24 percent—exceed 24k input tokens. By tailoring the hosted context window to match typical usage patterns, Cloudflare was able to offer a more accessible entry point for high-volume workflows. For applications requiring larger context windows, the standard Clef model retains its 64k capacity.
Clef Model Is Now Faster Through Serving Optimizations
In addition to expanding modalities and adjusting pricing, Cloudflare has boosted the processing speed of the standard Clef model through infrastructure-level enhancements on Workers AI. Because these improvements reside within the serving layer rather than the underlying model weights or architecture, no new model weights have been released.
A primary driver of this performance improvement was the migration to SGLang for model serving. Cloudflare collaborated with the SGLang development team to contribute pull requests integrating Clef support, which is slated for release in SGLang version 0.5.22. Developers interested in self-hosting Clef can access the updated launch commands through Cloudflare’s Hugging Face repository and the official SGLang documentation. Benchmarking data shows substantial speedups across various input sizes, with median processing times dropping significantly and delivering up to a twofold improvement for mid-sized inputs.
Real-World Applications Across Internal Teams
Internal experimentation at Cloudflare has demonstrated the practical utility of the Clef model family across diverse domains. By shifting detection and classification tasks directly into the model layer, development teams can integrate decision logic without needing to stand up specialized machine learning operations or curate large training datasets for domain-specific tuning.
Cloudflare teams have deployed Clef across several production and operational environments. The public GitHub documentation repository uses the model to automatically detect and close spam issues. The company’s Emdash content management system employs Clef to moderate plugin libraries for potential phishing threats. Additionally, the data loss prevention team utilizes the technology to scan data streams for sensitive personally identifiable information such as government identification numbers, while the threat intelligence group applies the models to assist in identifying malicious domains.
With the addition of multimodal capabilities, lowered pricing tiers, and accelerated processing speeds, the Clef model family offers developers a flexible toolset for modern AI infrastructure. Fully compatible with the Jev API and accessible via AI Gateway, the models allow developers to integrate advanced decision-making capabilities simply by updating their model identifiers.
Leave a Reply