Cloudflare open-sources Clef decision models and RL tuning

Cloudflare released Apache-2.0 Clef and Clef-flash multimodal decision models plus a customer RL fine-tuning platform.

Cloudflare open-sourced Clef, a 27-billion-parameter multimodal decision model post-trained from Qwen3.8-27B, and a smaller Clef-flash variant, both under Apache 2.0 and available on Hugging Face and Workers AI. The model accepts text, JSON, images, or video with a schema of typed questions and, via a joint schema head, returns one probability per allowed option in a single forward pass without free-form generation. Hugging Face coverage says the API is compatible with Jev and SystemOne and cites Decision Index 0.2.1 scores of 98.5 on BFCL, 69.2 nDCG@10 on ToolRet, and 91.9 accuracy on API-Bank; Cloudflare’s 2026-10-01 blog says Clef leads the Jev Decision Index but does not republish those figures or the SystemOne claim. The blog adds that Clef has a vision encoder and a 64k context window versus Jev’s 32k text-only design, that Clef-flash’s median latency is 38.8ms versus Jev’s 524.1ms, and that an internal threat-intel workflow classified a domain in 2.2 seconds versus 4.7 seconds for gpt-oss-120b. Cloudflare also introduced a reinforcement-learning fine-tuning platform for customer-specific adaptation. The reports do not conflict on license or the existence of both models; the blog omits the 27B size, Qwen3.8-27B base, and numeric Decision Index breakdown given in the earlier Hugging Face report.

• Cloudflare released Clef, a 27B multimodal model post-trained from Qwen3.8-27B under Apache-2.0, plus a smaller Clef-flash variant.
• Clef takes text, JSON, images, or video and a schema of typed questions, then a joint schema head returns one probability per allowed option in one forward pass without free-form generation.
• On Decision Index 0.2.1, the Hugging Face report gives Clef BFCL 98.5, ToolRet 69.2 nDCG@10, and API-Bank 91.9 accuracy; the Cloudflare blog says Clef leads the Jev Decision Index but does not repeat those scores.
• The Hugging Face report says the API is compatible with Jev and SystemOne; the blog says Clef adds a vision encoder and 64k context versus Jev’s 32k text-only design.
• Clef-flash median latency is 38.8ms versus Jev’s 524.1ms; in an internal threat-intel test Clef classified a domain in 2.2s versus 4.7s for gpt-oss-120b.
• Both models are on Hugging Face and Workers AI; Clef was trending #27 on Hugging Face as of 2026-09-30.
• Cloudflare also announced a reinforcement-learning fine-tuning product so customers can adapt Clef.