Cloudflare has announced a major expansion of its open-weight decision model ecosystem, introducing a new multimodal flagship model alongside significant price reductions and performance optimizations across its existing lineup. Just a week after the initial rollout of Clef and Clef-flash, the company is back with fresh developments designed to make AI-powered decision-making more versatile, affordable, and rapid for developers worldwide.
At the center of today’s announcements is the launch of Clef-omni. While the decision model space has remained largely text-centric since the debut of Jev from TypeSafe, Clef-omni breaks new ground by natively accepting audio and video inputs alongside text and images. Rather than forcing developers to construct complex, cascading pipelines that stitch together separate speech-to-text transcriptions or split audio and visual channels from video feeds, Clef-omni allows a single API call to handle decisions across all major modalities simultaneously.
The model represents a significant step toward artificial intelligence that can interpret the world through multiple human senses at once. Built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, Clef-omni processes text, imagery, audio, and video directly within a unified pipeline. By retaining the core comprehension backbone while stripping out text-to-speech generation components—since Clef models operate as decision engines rather than traditional large language models—the system eliminates the heavy overhead typically required for file transcription and captioning.
During execution, media elements map directly into a synchronized sequence where audio and video frames undergo joint processing. The system uses a two-stage attention routing mechanism to pull candidate values directly from internal embeddings, allowing valid options to gather evidence from any input modality. A built-in lexical grammar preserves option semantics, ensuring fast and schema-constrained scoring regardless of how the input is structured. Training followed the methodology established for the original Clef model, utilizing a frozen backbone, low-rank adapters, and a robust post-training combination of label-smoothed cross-entropy loss and Brier score calibration.
Performance metrics accompanying the launch show strong results across standard evaluation suites. Developers can now utilize Clef-omni to evaluate complex multi-modal workflows, such as checking equipment installations through a combination of product photos, running audio recordings, and operational video clips, all returned through a single structured request.
Alongside the introduction of Clef-omni, Cloudflare addressed developer feedback regarding cost by slashing prices for Clef-flash. The company optimized the model to position its pricing below that of competing alternatives like Jev, making it economical to scale agentic workflows across entire application stacks. To achieve this more accessible entry point, Cloudflare adjusted the hosted context window for Clef-flash to 24,000 tokens down from 64,000 tokens, noting that usage data demonstrated only a tiny fraction of requests—roughly 0.24 percent—ever exceeded the lower threshold. Developers who require larger context windows can still access the original 256k capability by self-hosting the model weights available on Hugging Face, while the standard Clef model maintains its 64k hosted context window.
Further enhancements were rolled out for the flagship Clef model, which received substantial speed improvements at the serving infrastructure layer. By migrating to SGLang for model serving—a contribution developed in collaboration with the SGLang team—Cloudflare achieved dramatic reductions in latency. Median execution speeds improved significantly across various input sizes, delivering roughly a 1.7-times speedup for smaller text inputs and doubling performance for medium-sized payloads, all without altering the underlying model weights.
Internal adoption at Cloudflare highlights the practical utility of the decision model family across diverse operational domains. Internal teams have deployed Clef to handle broad classification and detection tasks that previously would have required specialized machine learning infrastructure and dedicated training data corpora. Current production use cases include the public documentation repository automatically detecting and closing spam issues, the EmDash content management system scanning plugin libraries for phishing threats, data loss prevention pipelines screening for sensitive personally identifiable information such as government identification documents, and threat intelligence systems identifying malicious domains.
The expanded model family remains fully compatible with the Jev API and is accessible through AI Gateway, enabling developers to integrate the new capabilities simply by updating their model identifiers. With faster execution speeds, lower price points, and native multimodal support spanning text, images, audio, and video, Cloudflare continues to accelerate its iteration cycle in the machine learning space.
Leave a Reply