Just a week after making waves with the initial release of Clef and Clef-flash—its entry into the open-weight decision model space—Cloudflare is back with a broad expansion of its artificial intelligence lineup. The company has announced the immediate release of Clef-omni, a major new iteration that natively processes audio and video alongside traditional text and image inputs. Alongside the new multimodal capabilities, Cloudflare has slashed pricing for Clef-flash, positioning the model below competing offerings like Jev, while rolling out significant infrastructure-level speed enhancements for the flagship Clef model.
For Cloudflare, the rapid-fire cadence of these releases underscores a corporate culture centered on swift iteration and continuous engineering. The genesis of the entire Clef project reflects this philosophy, coming together in less than a week from conception to launch. Developers conceived the idea of building a dedicated decision model on a Friday evening, trained the architecture over the weekend, and deployed it to the public by Thursday. Despite the compressed timeline, the initial rollout delivered performant, high-quality, open-weight models to the broader developer community, setting the stage for even more ambitious capabilities moving forward.
Today’s momentum signals that Cloudflare’s footprint in the AI model ecosystem is expanding rapidly. By focusing on making models better, faster, cheaper, and more innovative, the company aims to embed intelligent decision-making seamlessly into everyday engineering and agentic workflows.
Clef-omni Takes Audio, Video, Image, and Text Input
The headline addition to the lineup is Clef-omni, a model designed to fundamentally change how decision models handle multimodal data. Since the introduction of TypeSafe’s Jev, the decision model paradigm has been largely restricted to text-only inputs. While the original Clef models added support for images and video frame arrays, Clef-omni dramatically broadens the scope by accepting audio files in WAV or MP3 formats and video files in MP4 or WebM formats, working concurrently with text and images.
This capability eliminates the need for complex, cascading engineering pipelines that previously required developers to set up separate speech-to-text transcription services or manually split audio and image channels from video feeds. Instead, developers can call a single model to evaluate and make decisions across any combination of data modalities simultaneously. This architecture brings developers one step closer to artificial intelligence that can interpret the world through the same multi-sensory channels humans experience—audio, visual, and textual communication—all within a unified request.
Under the hood, Clef-omni expands Cloudflare’s open-weight decision architecture by building upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. While the underlying foundation was already capable of processing text, imagery, audio, and video directly within a single pipeline, Cloudflare adopted the primary comprehension backbone while deliberately discarding the text-to-speech output components. During production execution, Clef-omni performs a rapid prefill pass across the complete payload, scoring all modalities and valid parameter options at the exact same time.
Because Clef models are specialized decision engines rather than traditional large language models, they skip output token generation entirely. This design choice removes the computational overhead typically associated with transcribing incoming audio files or generating captions for video elements. Instead, media elements map directly into a unified sequence, where video frames and audio streams are synchronized for joint processing. Clef-omni then extracts candidate values directly from internal embeddings using a two-stage attention routing mechanism. Every valid option gathers key evidence from the input—whether buried inside text snippets, visual elements, or audio streams—before cross-attending across the full context to compute exact confidence scores. A built-in lexical grammar preserves option semantics, ensuring fast, schema-constrained scoring across every input type.
Cloudflare trained the new model using the same methodology applied to the original Clef: freezing the Qwen3 backbone, training low-rank adapters, and employing a post-training approach that combines label-smoothed cross-entropy loss with Brier score calibration. The resulting model is highly resilient and engineered to succeed regardless of schema variations, field ordering, or prompt structures.
The performance profile of Clef-omni brings substantial upgrades to the family. Calibrated, schema-bound decisions can now effortlessly manage text, images, audio, and synchronized video in a single API call, while maintaining impressive latency metrics. Text-only decisions return in approximately 130 milliseconds at the median, image inputs take about 150 milliseconds, and audio clips are processed in a few hundred milliseconds. Even a full 21-second video clip complete with accompanying sound is scored in roughly 1.5 seconds through a single API invocation.
Comprehensive benchmarking shows that Clef-omni and its sibling models maintain robust performance across standard industry evaluations and TypeSafe evaluation suites, outperforming or competing closely with established alternatives across specialized workflows like invoice processing, customer service automation, security incident evaluation, and agent trace observability.
Clef-flash Model is Now Cheaper
Listening closely to developer feedback regarding the need for cost-effective decision models that can scale across extensive agentic workflows, Cloudflare has introduced aggressive price reductions for Clef-flash. Through careful optimizations, the company has lowered the price point of Clef-flash so that it is now more affordable than Jev, while preserving high performance and output quality.
To achieve this lower price point, Cloudflare made a specific architectural trade-off by adjusting the context window of Clef-flash. The hosted version of Clef-flash on Cloudflare’s infrastructure now features a 24k context window, down from the previously advertised 64k limit. However, the model weights available on Hugging Face remain untouched and retain full support for a 256k context window for developers choosing to self-host. Internal usage telemetry revealed that only a tiny fraction—0.24 percent—of requests ever exceeded 24k input tokens. Based on this real-world data, the company opted to reduce the hosted context window to establish a more accessible entry-level price point for high-volume workflows. The standard Clef model retains its full 64k context window for users with larger input requirements.
Clef Model is Now Faster
In addition to expanding modalities and lowering costs, Cloudflare has dramatically accelerated execution speeds for the flagship Clef model. These performance gains come entirely from optimizations implemented at the serving infrastructure layer on Workers AI, meaning no new model weights required release.
A primary driver of these speedups is the integration of SGLang for model serving. Collaborating directly with the SGLang development team, Cloudflare contributed pull requests to incorporate Clef into the ecosystem, a feature arriving in SGLang 0.5.22. Developers who prefer self-hosting can access the updated Hugging Face repository and SGLang cookbooks to deploy the optimized models on their own hardware.
How Are People Using Clef?
Since launch, internal teams at Cloudflare have actively experimented with Clef to handle classification and detection tasks across broad domains. Prior to the introduction of these decision models, deploying zero-shot classifiers often required organizations to stand up specialized machine learning teams, fine-tune smaller models for specific domains, and curate massive training corpora. Clef eliminates this friction, allowing developers to instantly parse complex input data with high precision through a simple model call.
Cloudflare teams are already leveraging Clef across multiple core infrastructure components. The public GitHub documentation repository uses the model to instantly detect and close spam issues. EmDash, Cloudflare’s content management system, relies on Clef to moderate plugin libraries and screen for potential phishing attempts. The data loss prevention team scans enterprise data for sensitive personally identifiable information such as government identification numbers, while threat intelligence teams use the models to identify and flag malicious domains.
Give It a Try!
The growing Clef family of models is fully compatible with the Jev API and accessible via AI Gateway, allowing developers to test the architecture simply by updating their model identifier. With newly added multimodal capabilities for audio and video, significantly reduced pricing, and accelerated infrastructure speeds, Cloudflare continues to expand the boundaries of what is possible in open-weight decision modeling.
Leave a Reply