Cloudflare is rapidly building out its presence in the artificial intelligence ecosystem. Just one week after debuting its initial open-weight decision models, Clef and Clef-flash, the company has announced a major expansion of the lineup. The centerpiece of this new release is Clef-omni, an advanced decision model capable of ingesting audio and video inputs alongside traditional text and images. Alongside this multimodal addition, Cloudflare has slashed the pricing for Clef-flash—making it more affordable than competing models like Jev—and implemented significant speed optimizations across the core Clef infrastructure.
The rapid pace of development highlights Cloudflare’s ongoing commitment to rapid iteration and continuous improvement. The entire concept of Clef originated in a matter of less than a week. The engineering team conceived the idea of entering the decision model space on a Friday evening, trained the model over the course of the weekend, and officially launched it the following Thursday. Delivering performant, high-quality open-weight models to the developer community on such a compressed timeline demonstrated the viability of the architecture, prompting the company to accelerate its roadmap and introduce further enhancements.
Clef-omni Takes Audio, Video, Image, and Text Input
The introduction of Clef-omni marks a significant shift in how decision models handle multimodal data. Since the introduction of TypeSafe’s Jev, decision models have largely remained restricted to text-only interactions. While Cloudflare’s initial Clef release added support for images and video frame arrays, Clef-omni natively ingests audio files in WAV or MP3 formats and video files in MP4 or WebM formats, working seamlessly alongside text and images.
This capability eliminates the need for complex, cascading pipelines that rely on separate speech-to-text transcriptions or require developers to split audio and visual channels from video feeds. Instead, applications can invoke 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 that humans experience daily.
Under the hood, Clef-omni expands Cloudflare’s open-weight decision architecture to natively manage complex multimodal workflows. Built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts (MoE) foundation, the model retains the underlying comprehension backbone while stripping out text-to-speech output components. During execution, Clef-omni performs a rapid prefill pass across the entire payload, scoring all modalities and valid parameter options at the same time.
Because Clef models function as decision-makers rather than traditional Large Language Models (LLMs), they bypass output token generation entirely. This removes the performance overhead associated with transcribing or captioning incoming media files. Instead, media elements map directly into a unified sequence where video and audio sync with visual frames for joint processing. Clef-omni extracts candidate values straight from internal embeddings using a two-stage attention routing mechanism, where every valid option gathers key evidence from text, visual elements, or audio streams before cross-attending across the full context to determine confidence scores. A built-in lexical grammar preserves option semantics, delivering schema-constrained scoring across every input type.
The training methodology mirrors that of the original Clef model, utilizing frozen Qwen3 backbones, low-rank adapters (LoRA), and a post-training approach that combines label-smoothed cross-entropy loss with Brier score calibration. This design ensures robust performance capable of handling schema variations, field ordering differences, and diverse prompt structures. Consequently, developers can execute calibrated, schema-bound decisions spanning text, images, audio, and synchronized video through a single API call with minimal latency. Text-only decisions return in approximately 130 milliseconds at the median, image inputs in about 150 milliseconds, and audio clips in a few hundred milliseconds, while a full 21-second video clip with sound is scored in roughly 1.5 seconds.
Extensive benchmarking demonstrates strong performance across standard evaluations and TypeSafe evaluation frameworks, illustrating the effectiveness of the MoE architecture and newly integrated modalities.
Clef-flash Model is Now Cheaper
Responding directly to developer feedback requesting a decision model affordable enough for integration into high-volume workflows, Cloudflare has optimized Clef-flash to lower its price point. The updated pricing structure positions Clef-flash as more affordable than Jev, encouraging widespread adoption across extensive agentic applications.
To achieve this price reduction, Cloudflare made a strategic adjustment to the hosted model’s context window. The hosted version of Clef-flash now features a 24k context window, down from the previously advertised 64k. However, the model weights available on Hugging Face remain untouched and retain native support for a 256k context window for developers choosing to self-host. Internal usage telemetry indicated that only 0.24 percent of requests exceed 24k input tokens, justifying the trade-off to provide a more accessible entry point. The standard Clef model retains its 64k context window for workloads requiring larger capacities.
Clef Model is Now Faster
In addition to expanding modality support and lowering costs, Cloudflare has accelerated the core Clef model through infrastructure-level optimizations on Workers AI rather than alterations to the model weights or architecture. A key driver of these speed improvements is the migration to SGLang for model serving. Following collaborative pull requests with the SGLang team, the optimizations are incorporated into SGLang 0.5.22, with updated launch commands provided in Cloudflare’s Hugging Face repository and SGLang documentation for self-hosting enthusiasts.
These infrastructure enhancements yield substantial speedups. For inputs around 800 tokens, median processing times dropped from 262 milliseconds to 152 milliseconds. Workloads involving approximately 3,400 tokens saw median response times cut from 616 milliseconds to 305 milliseconds, achieving a twofold speedup. Larger payloads of around 16,000 tokens improved from 2,721 milliseconds down to 1,635 milliseconds.
How are People Using Clef?
Internal teams at Cloudflare have actively experimented with Clef for generalized domain classification and detection tasks, demonstrating the practical utility of embedding decision capabilities directly into the model layer. Rather than requiring specialized machine learning teams, extensive domain training, or custom data corpuses to build zero-shot classifiers, developers can utilize Clef to instantly parse input data with high precision.
Internal deployments highlight the versatility of the model family. Cloudflare’s public GitHub documentation repository uses Clef to automatically detect and close spam issues. The EmDash content management system relies on the models to scan and moderate plugin libraries for potential phishing threats. Meanwhile, the data loss prevention team uses Clef to scan incoming data for sensitive personally identifiable information such as government identification documents, and the threat intelligence team employs the models to assist in identifying malicious domains.
Building on the positive reception of its initial model releases, Cloudflare’s latest updates make its decision architecture faster, more affordable, and capable of processing complex multimodal inputs. Fully compatible with the Jev API and accessible via AI Gateway, the updated Clef models provide developers with an expanded toolkit for intelligent automation and classification workflows.
Leave a Reply