Cloudflare has released Clef and Clef-flash, its first AI models trained in-house. The launch was announced on October 1. Unlike chatbots, these are decision models. They answer set questions with probabilities instead of generating text. Both are open-weight under the Apache 2.0 licence. Developers can download them from Hugging Face. Clef has 27 billion parameters. Clef-flash is a smaller 9 billion parameter version built for speed. Both use a Qwen backbone and a 64,000-token context window. They run on Cloudflare’s Workers AI platform. The models are designed for AI agents that need fast, structured decisions. Examples include routing support tickets or flagging fraud. Cloudflare says Clef beats rivals on 7 of 10 decision benchmarks. It is the company’s biggest step yet into AI model development. Developers have welcomed the open weights, with early tests praising the models’ speed on routing tasks. More model sizes may follow later this year. The release marks a new chapter for the company.