Following last week’s debut of Clef and Clef-flash, Cloudflare is significantly expanding its footprint in the decision model ecosystem with the official release of Clef-omni. The new multimodal model introduces native support for audio and video inputs alongside the text and image processing capabilities established by its predecessors. Alongside this major architectural milestone, the infrastructure and security giant has slashed prices for Clef-flash—making it more affordable than rival models like Jev—while rolling out major serving infrastructure optimizations that dramatically accelerate Clef processing speeds across the board.
Although Cloudflare’s journey into the artificial intelligence model space is still in its earliest stages, rapid innovation and iteration have long been central to the company’s operational ethos. The entire trajectory of the Clef initiative underscores this fast-paced development cycle, coming together over the course of less than a week. The engineering team conceived the idea of building a dedicated decision model framework on a Friday evening, completed the training phase over the weekend, and successfully launched the models by the following Thursday. Despite such an extraordinarily compressed timeline, the team delivered performant, high-quality, open-weight models designed to serve the broader developer community.
Building on that initial momentum, the introduction of Clef-omni marks a paradigm shift for decision models. Since the market debut of Jev from TypeSafe, decision models have remained overwhelmingly text-centric. While the original Clef release introduced support for static images and video frame arrays, Clef-omni natively ingests audio files in WAV or MP3 formats alongside MP4 or WebM video streams, directly integrated with text and imagery within a single unified architecture.
This capability eliminates the traditional engineering overhead of setting up cumbersome, cascading pipelines where developers must first deploy speech-to-text transcription models or manually split audio and image channels from incoming video feeds. Instead, developers can now call a single model to evaluate complex scenarios and execute decisions across any modality simultaneously. This unified approach brings the industry one step closer to artificial intelligence systems capable of interacting with the physical world through the same multimodal spectrum—audio, visual, and textual communication—experienced by humans every day.
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 engineered to process text, imagery, audio, and video directly within a single pipeline, Cloudflare adapted the primary comprehension backbone while deliberately discarding traditional text-to-speech output components. In a production environment, Clef-omni executes a rapid prefill pass across the complete payload, scoring all modalities and valid parameter options simultaneously.
Because decision models function differently than standard large language models and do not generate output text tokens, Clef avoids the computational overhead typically required for transcribing or captioning incoming files. Media elements map directly into a unified sequence where video and audio are synchronized with visual frames for joint processing. Clef-omni subsequently extracts candidate values straight from internal embeddings using a sophisticated two-stage attention routing mechanism. Every valid option gathers essential evidence from the input—whether those indicators are embedded in 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 and schema-constrained scoring across every input type.
The training methodology mirrors the approach used for the original Clef models: freezing the Qwen3 backbone, training low-rank adapters, and applying a rigorous post-training approach that combines label-smoothed cross-entropy loss with Brier score calibration. This engineering rigor results in a resilient model designed to maintain high performance despite variations in schemas, field ordering, and prompt structures.
Performance benchmarks confirm that Clef-omni delivers rapid response times across diverse inputs. Text-only decisions resolve in a median time of approximately 130 milliseconds, image inputs process in roughly 150 milliseconds, and standalone audio clips return in a few hundred milliseconds. Even a full 21-second video clip accompanied by audio is evaluated and scored in about 1.5 seconds through a single API call. In benchmark evaluations comparing the model against industry alternatives such as Jev across diverse tasks ranging from tool retrieval and API utilization to customer service and banking intent classification, the Clef family demonstrates highly competitive results.
Clef-flash Model Becomes More Affordable
Responding directly to developer feedback calling for decision models economical enough to integrate into high-volume workflows at scale, Cloudflare has introduced aggressive price reductions for Clef-flash. Following a series of targeted internal optimizations, Clef-flash is now priced lower than competing options like Jev, creating a strong economic incentive for organizations to deploy agentic workflows across their entire operational stacks.
To achieve this more accessible price point, Cloudflare made a deliberate architectural trade-off by reducing the hosted version’s context window. The managed version of Clef-flash now features a 24k context window, down from the previously advertised 64k limit. However, the model weights available on Hugging Face remain entirely untouched and retain their native training to support a 256k context window for developers choosing to self-host the model.
Internal usage telemetry heavily informed this decision, revealing that a mere 0.24 percent of requests ever exceed 24k input tokens. By adjusting the hosted context window to accommodate the vast majority of standard use cases, Cloudflare successfully lowered the barrier to entry. For applications requiring larger contexts, the standard Clef model maintains its 64k window as a robust alternative.
Clef Model Achieves Significant Speed Improvements
In addition to expanding model modalities and reducing pricing, Cloudflare has significantly accelerated the processing speed of the standard Clef model. These performance gains stem from targeted optimizations applied at the serving infrastructure layer rather than structural alterations to the underlying model weights or architecture.
A primary driver of this speedup was the transition to SGLang for model serving. Cloudflare collaborated closely with the SGLang development team to contribute pull requests integrating Clef support, which will officially debut in SGLang version 0.5.22. Developers preferring to self-host Clef can access the updated SGLang launch commands through Cloudflare’s updated Hugging Face repository and the official SGLang documentation cookbooks.
These infrastructure-level enhancements have yielded dramatic performance improvements across varying input sizes. For inputs containing approximately 800 tokens, median processing times dropped from 262 milliseconds down to 152 milliseconds, achieving a 1.7x speedup. For mid-sized payloads around 3,400 tokens, median latency was cut in half from 616 milliseconds to 305 milliseconds, representing a 2.0x performance boost. Larger payloads containing approximately 16,000 tokens saw median processing times decrease from 2,721 milliseconds to 1,635 milliseconds, maintaining a 1.7x speedup.
Real-World Adoption Across Internal Teams
Internal teams at Cloudflare have already begun experimenting with Clef as a versatile classification model operating across broad domains. Historically, deploying zero-shot classifiers for specialized tasks required organizations to stand up dedicated machine learning engineering teams, curate extensive data corpuses, and fine-tune smaller models to achieve domain-specific accuracy. Clef abstracts this complexity into the model layer, allowing engineers to instantly parse through diverse input data with high precision via simple API calls.
Several internal use cases highlight the practical utility of the Clef model family. Cloudflare’s public GitHub documentation repository utilizes Clef to automatically detect and close spam issues. The company’s EmDash content management system employs the models to moderate plugin libraries and screen for potential phishing threats. Meanwhile, data loss prevention teams scan structured and unstructured data for sensitive personally identifiable information such as government identification numbers, and threat intelligence teams leverage the models to flag malicious domains.
Available as a fully Jev-API compatible endpoint accessible via AI Gateway, the Clef model family requires minimal migration effort for developers—often requiring nothing more than updating the model identifier. As Cloudflare continues to expand its open-weight artificial intelligence portfolio with multimodal capabilities, improved cost efficiency, and accelerated serving speeds, the company invites the developer community to test the models and share feedback on their implementations.
Leave a Reply