Cloudflare Clef and Clef-flash Open-Source Decision Models

Features

• ✓ Cloudflare says Clef is a decision model that returns typed answers with probabilities (such as choice and score) rather than generating text, letting agents make programmatic decisions without necessarily keeping a human in the loop, for example judging whether a support message is urgent and which team should handle it
• ✓ Cloudflare says Clef has a vision encoder so it can classify visual content such as images, and has a 64k context window (Cloudflare compares this with Jev's 32k) to hold more input state
• ✓ Cloudflare says Clef and Clef-flash are fully Jev-API compatible so they can be swapped in easily; the larger Clef targets precision while Clef-flash targets latency-critical decisions (Cloudflare reports median latency of about 209.3ms and 38.8ms respectively)
• ✓ Cloudflare says both models are open-sourced on Hugging Face under Apache 2.0 for local experimentation and are hosted on Workers AI edge GPUs to reduce network latency
• ✓ Cloudflare is also debuting its reinforcement-learning (RL) fine-tuning product so customers can fine-tune Clef with their own data; Cloudflare says it does not read, store or train on customer requests or responses unless they choose to use the fine-tuning product

How to Use

1. Open the Workers AI Clef model page in Cloudflare's developer documentation, or download the open weights from Hugging Face (huggingface.co/Cloudflare/clef)
2. Call Clef through the Workers AI REST API; Cloudflare's example uses curl to POST to /ai/run/@cf/cloudflare/clef with your account ID and auth token
3. Pass the input as state and define questions with types such as choice and score; the model returns typed answers with probabilities
4. For specialized needs, use Cloudflare's new RL product to fine-tune Clef on your own data and redeploy the fine-tuned model on Workers AI

FAQ

Official Source