Following the initial release of its open-weight decision models last week, Cloudflare has announced a major expansion of the lineup with the launch of Clef-omni, alongside significant price reductions for Clef-flash and infrastructure-level performance boosts for the flagship Clef model. The rapid rollout highlights the company’s aggressive iteration cycle in the artificial intelligence space, developing from concept to launch within compressed timeframes while continuing to build out specialized tools for community and enterprise developers alike.
The model game remains in its early stages for Cloudflare, but the engineering team has leaned heavily into its foundational principles of innovation and rapid iteration. The original conception of Clef materialized in less than a week, originating from a Friday evening brainstorming session, moving through weekend training, and culminating in a Thursday launch. Delivering performant, high-quality open-weight models on such tight schedules has set the stage for further developments, culminating in today’s introduction of new features, modalities, and performance enhancements across the entire Clef family.
Clef-Omni Takes Audio, Video, Image, and Text Input
The headline addition to the portfolio is Clef-omni, which natively processes audio, video, image, and text inputs within a single workflow. This capability marks a shift for decision models, which have remained largely text-centric since the introduction of competing options such as Jev from TypeSafe. While the original Clef models introduced support for static images and video frame arrays, Clef-omni extends this capability by ingesting audio formats like WAV or MP3 alongside video files such as MP4 or WebM, blending them directly with accompanying text and visual data.
By unifying these inputs, developers can eliminate complex cascading pipelines that traditionally rely on separate speech-to-text transcription services or require manual splitting of audio and video channels. Instead, a single API call can execute decisions across multiple modalities simultaneously. This architecture brings automated decision-making one step closer to interacting with real-world scenarios through simultaneous audio, visual, and textual communication.
Under the hood, Clef-omni builds upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. Cloudflare adopts the primary comprehension backbone of this model while removing text-to-speech output components, ensuring the system operates purely as a decision engine rather than a generative chatbot. During execution, the model performs a rapid prefill pass across the complete payload, scoring all modalities and valid parameters concurrently.
Because Clef models bypass output token generation entirely, they avoid the computational overhead associated with transcribing or captioning incoming files. Media elements map directly into a unified sequence where video and audio sync with visual frames for joint processing. Candidate values are then pulled directly from internal embeddings using a two-stage attention routing mechanism, where every valid option gathers key evidence from text snippets, visual elements, or audio streams. A built-in lexical grammar preserves option semantics, delivering schema-constrained scoring across every input type.
Training utilizes the same techniques established for the original Clef model, featuring a frozen Qwen3 backbone combined with low-rank adapters and a post-training methodology that merges label-smoothed cross-entropy loss with Brier score calibration. The resulting architecture maintains resilience against schema variations, field ordering shifts, and prompt structures. Consequently, calibrated and schema-bound decisions handle text, images, audio, and synchronized video within a single API call while returning results quickly. Median processing times span approximately 130 milliseconds for text-only inputs, 150 milliseconds for images, and a few hundred milliseconds for audio clips, while scoring a full 21-second video clip with sound takes roughly 1.5 seconds.
Clef-Flash Model Is Now Cheaper
In response to developer feedback calling for a decision model affordable enough for widespread integration across automated workflows, Cloudflare has implemented optimizations to lower the pricing of Clef-flash. The adjustments position Clef-flash below competing alternatives like Jev in cost while preserving performance and quality standards.
To achieve this pricing structure, Cloudflare made a deliberate trade-off regarding the context window of Clef-flash. The hosted version of Clef-flash now features a 24-kilobyte context window, down from the previously advertised 64-kilobyte limit. However, the model weights available on Hugging Face remain untouched and retain support for a 256-kilobyte context window for developers choosing to self-host. Internal usage telemetry indicated that only a fraction of a percent of requests—specifically 0.24 percent—exceeded 24,000 input tokens, justifying the adjustment to provide a more accessible entry point. The standard Clef model retains its 64-kilobyte context window for applications requiring larger capacity.
Clef Model Is Now Faster
Alongside the new model and pricing updates, Cloudflare has accelerated the flagship Clef model through optimizations implemented directly within the serving infrastructure layer on Workers AI, requiring no changes to the underlying model weights or architecture.
A primary driver of these speed gains is the adoption of SGLang for model serving. Cloudflare collaborated with the SGLang team to integrate Clef support, which is slated for release in SGLang version 0.5.22. Developers interested in self-hosting can access the updated launch commands through Cloudflare’s Hugging Face repository and the official SGLang documentation cookbooks.
Performance metrics demonstrate notable improvements across varying input sizes. For inputs around 800 tokens, median processing times dropped from 262 milliseconds to 152 milliseconds, achieving a 1.7-times speedup. Inputs around 3,400 tokens saw median speeds improve from 616 milliseconds down to 305 milliseconds, representing a twofold increase in speed. Larger workloads consisting of approximately 16,000 tokens improved from a median of 2,721 milliseconds down to 1,635 milliseconds, maintaining a 1.7-times speedup.
How Are People Using Clef?
Internal teams at Cloudflare have actively experimented with Clef for generalized domain classification and detection tasks. Moving detection capabilities directly into the model layer allows developers to incorporate decision logic without needing to stand up specialized machine learning teams or curate extensive domain-specific training datasets.
Several internal applications demonstrate the practical utility of the model family. The public GitHub documentation repository utilizes Clef to automatically detect and close spam issues. The EmDash content management system applies the models to moderate plugin libraries for potential phishing threats. Meanwhile, Cloudflare’s data loss prevention team uses the technology to scan data streams for personally identifiable information, such as government identification numbers, and the threat intelligence team deploys Clef to identify malicious domains.
Give It A Try!
The expanded Clef model family, featuring native multimodal inputs, reduced pricing, and enhanced execution speeds, is fully compatible with the Jev API and accessible via AI Gateway. Developers can begin testing the updated models by updating their model identifiers, with complete technical documentation available through Cloudflare’s developer portal.
Leave a Reply