Following the recent introduction of its open-weight decision models, Cloudflare is expanding its artificial intelligence offerings with the release of Clef-omni, a multimodal model capable of processing audio, video, images, and text simultaneously. Alongside this major capability upgrade, the company has announced aggressive price reductions for Clef-flash—making it more affordable than rival models like TypeSafe’s Jev—and introduced significant infrastructure speed improvements for the flagship Clef model.
The rapid pace of development underscores Cloudflare’s iterative approach to engineering. Conceived and trained over the course of a single weekend, the original Clef models demonstrated that performant, specialized decision frameworks could be deployed quickly to the open-source community. Building on that momentum, the company’s latest release aims to bring comprehensive, real-world data processing to developers without the heavy computational overhead typically associated with traditional large language models.
Clef-Omni Takes Audio, Video, Image, and Text Input
The introduction of Clef-omni marks a notable shift in how decision models handle multimodal inputs. While standard decision models have remained largely text-centric since the debut of TypeSafe’s Jev, and earlier iterations of Clef supported images and video frame arrays, Clef-omni natively ingests audio files in WAV or MP3 formats alongside MP4 or WebM video streams, paired with traditional text and images.
This architecture eliminates the need for complex, cascading pipelines that rely on separate speech-to-text transcription services or require developers to manually split audio and visual channels from video feeds. Instead, applications can issue a single API call to evaluate decisions across multiple modalities at once. By processing audio, visual, and textual inputs simultaneously within a unified framework, the model moves a step closer to interacting with complex environments much like human operators do.
Built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, Clef-omni adapts the core comprehension backbone while omitting traditional text-to-speech output components. During execution, the model performs a rapid prefill pass over the entire payload, scoring all input modalities and potential parameters concurrently. Because Clef-omni functions as a specialized decision model rather than a traditional large language model, it skips output token generation entirely. Media elements map directly into a unified sequence where video frames and audio tracks are synchronized for joint processing, allowing the system to extract candidate values straight from internal embeddings using two-stage attention routing.
Clef-Flash Model Is Now Cheaper
Responding directly to developer feedback regarding the cost of scaling automated workflows, Cloudflare has optimized its infrastructure to offer Clef-flash at a significantly lower price point. The model is now positioned as a more economical choice than competing offerings such as Jev, lowering the financial barrier for organizations seeking to embed decision intelligence across high-volume agentic pipelines.
Achieving this reduced pricing structure required a strategic adjustment to the hosted version’s context window. Clef-flash now features a hosted context window of 24,000 tokens, down from the previously advertised 64,000 tokens. However, the open-weights version available on Hugging Face remains entirely untouched, retaining its native support for a 256k context window for developers who choose to self-host.
Usage data played a decisive role in the decision, revealing that only a minute fraction of requests—approximately 0.24 percent—exceed 24,000 input tokens. For organizations requiring larger context capacities, Cloudflare continues to recommend the standard Clef model, which maintains its original 64k context window.
Clef Model Is Now Faster
Performance enhancements have also arrived for the flagship Clef model through infrastructure-level optimizations on Workers AI, requiring no alterations to the underlying model weights. By transitioning to SGLang for model serving—a collaboration that yielded pull requests incorporated into the SGLang project—Cloudflare has drastically reduced latency across various input sizes.
Benchmarks shared by the company highlight substantial speedups across text and token inputs. For payloads averaging around 800 tokens, median latency dropped from 262 milliseconds to 152 milliseconds. Inputs scaling to approximately 3,400 tokens saw response times cut in half, dropping from 616 milliseconds to 305 milliseconds. Even large payloads reaching 16,000 tokens experienced marked improvements, with median response times falling from over 2.7 seconds down to approximately 1.6 seconds. Developers opting to self-host can access updated launch commands and deployment guides directly through the platform’s updated repositories and serving documentation.
How Are People Using Clef?
Inside Cloudflare, internal engineering teams have begun deploying the Clef family across a diverse array of production use cases, shifting complex detection and classification tasks directly into the model layer. Previously, implementing zero-shot classification often required specialized machine learning infrastructure and extensive domain-specific training data. The streamlined architecture of Clef allows systems to parse incoming data instantly with high precision.
Real-world deployments within the organization reflect the versatility of the model family. The public GitHub documentation repository utilizes Clef to automatically detect and close spam issues. The company’s content management system employs the models to screen plugin libraries for potential phishing threats. Meanwhile, data loss prevention teams scan information streams for sensitive personally identifiable information such as government identification numbers, and threat intelligence groups leverage the models to identify malicious domains.
Available through AI Gateway with full compatibility for Jev-API structures, the expanded Clef model family offers developers a straightforward path to integrating multimodal decision capabilities into existing software architectures.
Leave a Reply