# Introducing Clef-omni: Cloudflare’s Multimodal Decision Model with Major Price and Speed Upgrades
Cloudflare has made significant strides in the decision model space with a series of updates to its Clef family of models. These open-weight models, designed for classification, detection, and decision-making tasks, have now expanded into new territory with multimodal capabilities, more affordable pricing, and faster inference speeds. The entire Clef lineup has undergone meaningful improvements that make decision models more accessible and versatile than ever before.
## Clef-omni: A New Era of Multimodal Decision-Making
The standout announcement is Clef-omni, a new model that natively accepts audio, video, image, and text input all within a single API call. This marks a significant shift from traditional decision models, which have historically been limited to text-only processing.
What makes Clef-omni particularly powerful is its ability to process multiple media types simultaneously. Rather than requiring developers to set up cascading pipelines where audio must first be transcribed to text, or where video frames and audio channels must be split and processed separately, Clef-omni handles everything in one unified pass. Whether a user needs to analyze a photograph, listen to an audio clip, review a video, or read a text prompt — all at once — the model delivers a single decision.
The architecture behind Clef-omni builds upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. The team adopted the primary comprehension backbone while removing text-to-speech output components, creating a model optimized for decision scoring rather than text generation. A two-stage attention routing mechanism allows the model to gather evidence from any modality — whether buried in text snippets, visual elements, or audio streams — and compute confidence scores across all candidate options simultaneously.
Performance is impressive given the complexity. Text-only decisions return in approximately 130 milliseconds at the median, image inputs take about 150 milliseconds, and audio clips process in a few hundred milliseconds. Even a full 21-second video clip with accompanying sound is scored in around 1.5 seconds, all within a single API call.
## Pricing Cuts Make Clef-flash More Accessible
In response to community feedback, Cloudflare has significantly reduced the price of Clef-flash. The model is now available at $0.038 per million input tokens, down from $0.09 — making it cheaper than the previously dominant Jev model, which sits at a comparable price point.
The new pricing structure across the Clef family is as follows:
– **Clef-flash**: $0.038 per million input tokens
– **Clef**: $0.24 per million input tokens
– **Clef-omni**: $0.15 per million input tokens
A trade-off accompanies this price reduction: the hosted version of Clef-flash now operates with a context window of 24,000 tokens, down from the previously advertised 64,000 tokens. However, the model weights available on Hugging Face remain unchanged and support up to 256,000 tokens for those who choose to self-host. Cloudflare noted that only 0.24% of requests exceeded 24,000 tokens, making this a practical adjustment to broaden accessibility. Users with larger context requirements are encouraged to use the full Clef model, which retains its 64,000-token window.
## Speed Improvements Across the Clef Family
The hosted Clef model has also received notable speed optimizations. Median inference times have been reduced significantly across various input sizes:
– For inputs of approximately 800 tokens, response times improved from 262 milliseconds median to 152 milliseconds — a 1.7× speedup.
– For inputs of approximately 3,400 tokens, the median dropped from 616 milliseconds to 305 milliseconds — a 2.0× speedup.
– For inputs of approximately 16,000 tokens, the median fell from 2,721 milliseconds to 1,635 milliseconds — a 1.7× speedup.
These improvements were achieved primarily through infrastructure-level optimizations rather than changes to model weights. Cloudflare migrated to SGLang as the serving framework for Clef, contributing pull requests back to the SGLang project (PR #42721) which will be included in SGLang version 0.5.22. Self-hosted users can take advantage of updated launch commands available on the Clef Hugging Face repository and SGLang cookbooks.
## Benchmark Performance
The Clef family continues to demonstrate strong benchmark results. Clef-omni achieved a score of 98.2 on BFCL case exact and 94.8 on BANKING77 macro-F1, outperforming Jev on the latter metric significantly. Clef-flash leads in BFCL case exact at 98.76 and API-Bank accuracy at 93.11. Across the TypeSafe evals for workflow tasks such as invoice processing and customer service, the models consistently perform competitively against established alternatives.
## Real-World Adoption
Teams within Cloudflare have already begun integrating Clef into production workflows. The company’s public GitHub documentation repository uses Clef to detect and automatically close spam issues. EmDash, Cloudflare’s content management system, employs Clef to moderate plugin libraries against phishing attempts. The data loss prevention team leverages the model to scan for personally identifiable information, including government identification numbers. Additionally, the threat intelligence team uses Clef to identify malicious domains.
This adoption highlights a key advantage of decision models like Clef: the capability to perform detections and classifications has moved directly into the model layer, eliminating the need for specialized machine learning teams to train domain-specific classifiers from scratch. Any developer can now call a model and instantly parse input data with precision, regardless of the domain.
## Getting Started
Clef is fully compatible with the Jev API and is available through AI Gateway. Developers can start using Clef today simply by changing the model ID in their existing setup. All models are open-weight and available on Hugging Face, with comprehensive developer documentation provided by Cloudflare.
The Clef family represents Cloudflare’s commitment to making powerful, multimodal decision-making tools accessible to the broader developer community. With new modalities, lower prices, and faster speeds, the company continues to iterate rapidly on what was initially a weekend project turned into a fully shipping model family.
## Frequently Asked Questions
**What is a decision model?**
A decision model is a type of AI model designed to classify inputs, detect patterns, and make structured decisions — rather than generating free-form text like a large language model. Decision models are commonly used for tasks such as intent classification, spam detection, and domain categorization.
**What makes Clef-omni different from Clef and Clef-flash?**
Clef-omni is the first model in the Clef family to natively accept multimodal input, including audio, video, images, and text. Previous Clef models handled text and image inputs, while Clef-omni expands this to also process audio files and video clips directly.
**Can I self-host Clef models?**
Yes, all Clef models are open-weight and available on Hugging Face for self-hosting. The Clef-flash model weights support up to 256,000 tokens in context, even though the hosted version is limited to 24,000 tokens.
**Why was the Clef-flash context window reduced?**
The context window was reduced from 64,000 to 24,000 tokens in the hosted version to enable significantly lower pricing. Cloudflare’s usage data showed that only 0.24% of requests exceeded the 24,000-token threshold, making this a practical and impactful optimization.
**What benchmarks does Clef-omni excel in?**
Clef-omni leads in BANKING77 macro-F1 (94.8) and CLINC150+OOS macro-F1 (97.7), demonstrating strong performance in financial intent classification and out-of-scope detection tasks.
**How does Clef compare to Jev in pricing?**
Clef-flash is now priced at $0.038 per million input tokens, which is cheaper than Jev. Clef-omni launches at $0.15 per million input tokens, placing it in a competitive mid-range tier.
**What frameworks are used to serve Clef models?**
Cloudflare uses SGLang for serving Clef models on Workers AI. The integration has been contributed upstream to SGLang version 0.5.22.
## Conclusion
The Clef model family has evolved rapidly from a weekend experiment into a robust, multimodal decision-making toolkit. The introduction of Clef-omni brings audio and video processing into a single unified model, the price reduction of Clef-flash makes decision models viable for any scale of project, and infrastructure-level speed improvements ensure that these models remain performant under real-world conditions. As Cloudflare continues to iterate on this family, the open-weight approach ensures that developers everywhere can access, modify, and deploy these models in their own workflows. The decision model landscape is changing, and Clef is at the forefront of that shift.
Thank you for reading
