Cloudflare/clef
Cloudflare's Clef is a 27 billion parameter multimodal model, post-trained from Qwen/Qwen3.8-27B, designed to convert states and schemas of typed questions into decisions. It processes inputs across text, JSON, images, or video, and outputs probabilities for all allowed options of each question in a single forward pass. This model specializes in structured decision-making, eliminating free-form text generation and output parsing, making it ideal for automated decision systems.
Loading preview...
Clef: A Multimodal Decision Model
Clef is a 27 billion parameter multimodal model developed by Cloudflare, specifically engineered for structured decision-making. Post-trained from Qwen/Qwen3.8-27B, it takes a given state (which can be text, JSON, images, or video) and a schema of typed questions, then outputs probabilities for every allowed option of each question. This model is unique in its approach, focusing on direct decision outputs rather than free-form text generation, and is fully compatible with Jev and SystemOne APIs.
Key Capabilities
- Multimodal Input: Processes information from text, JSON, images, and video to inform decisions.
- Structured Decision Output: Provides a probability for each option within a defined schema, eliminating the need for output parsing.
- Efficient Inference: Performs a single forward pass to generate all decision probabilities.
- Specialized Architecture: Utilizes a Qwen/Qwen3.8-27B backbone with a vision encoder, augmented by a joint schema head that routes evidence and scores options.
- API Compatibility: Seamlessly integrates with Jev and SystemOne APIs for streamlined workflow automation.
Performance Highlights
Clef demonstrates strong performance across various benchmarks on the Decision Index leaderboard. It shows competitive or leading results in tasks such as ToolRet (69.2% nDCG@10), API-Bank (91.9% accuracy), BANKING77 (94.2% macro-F1), and CRUXEval (86.7% accuracy). In workflow evaluations, Clef excels in "Invoice processing" with 64.7% exact actions and "Security incidents" with 62.9% exact actions, often outperforming its smaller variant, Clef-Flash, and other models like Jev.
Use Cases
Clef is particularly well-suited for applications requiring automated, structured decision-making based on diverse inputs, such as:
- Invoice Processing: Automatically determining invoice status or actions.
- Customer Service Routing: Directing inquiries to the correct department and assessing urgency.
- Security Incident Response: Evaluating the nature and severity of security events.
- Any system requiring probabilistic choices from structured data.