Cloudflare has launched its first internally developed open-weight AI models, Clef and Clef-flash, on its Workers AI platform. These models are designed for fast, structured decision-making tasks, such as request categorization or task routing within agentic workflows.
The Clef model features 27 billion parameters based on Qwen3.8-27B, a 64k-token context window, and support for text, JSON, images, and video. The lighter Clef-flash model uses 9 billion parameters based on Qwen3.5-9B and is optimized for latency-critical applications.
Both models incorporate a vision encoder capable of processing up to 4 images per request. They are available on Hugging Face under the Apache 2.0 license and run on Cloudflare’s edge GPU infrastructure to minimize data travel distance and latency.
Pricing for the models is set at $0.24 per million input tokens for Clef and $0.09 per million input tokens for Clef-flash, with no charges for output tokens. Both models are compatible with the Jev-API to facilitate integration.
In performance tests using the BANKING77 benchmark, Clef achieved a macro-F1 score of 94.20. Clef-flash demonstrated a median latency of 38.8 ms. Cloudflare is also introducing a reinforcement learning fine-tuning platform to support these models.
