Cloudflare has introduced Clef and Clef-flash, two open-weight decision models designed for specific question answering rather than general text generation. These models, built upon Qwen backbones and compatible with the TypeSafe AI Jev API, return typed probabilities for predefined questions about an input state, including text, images, and video. While Clef and Clef-flash demonstrate superior performance on certain benchmarks like Banking77 and CLINC150+OOS, they lag behind Jev in knowledge-intensive tasks such as GPQA Diamond and MMLU-Pro. Cloudflare is also offering a reinforcement learning service to fine-tune these models on private data. AI
IMPACT These models offer a new paradigm for structured decision-making with LLMs, potentially improving efficiency and reliability in specific applications.
RANK_REASON Cloudflare released new open-weight models. [lever_c_demoted from frontier_release: ic=2 ai=1.0]
- Banking77
- Clef-flash
- CLINC150+OOS
- Cloudflare
- GPQA Diamond
- GPT-OSS 120B
- MMLU-Pro
- Qwen3.5:9b
- Qwen3.8-27B
- TypeSafe AI
- Workers AI
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →