- Published on
Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text
- Authors
- Name
- DeepHorus
Cloudflare has released Clef and Clef-flash, the first models trained by its Workers AI team, as open-weight models for structured decisions. Instead of generating conversational text, they accept an input state and typed-question schema, then return probabilities for every allowed answer. Both are released under Apache 2.0 and support TypeSafe AIās Jev API, so migration mainly requires changing the endpoint and model name. The launch signals Cloudflareās push into task-specific inference infrastructure rather than another general-purpose chatbot.
The design targets a production problem: free-form output must be parsed, validated, and mapped to application actions. Clef supports noul for yes/no decisions, choice for named options, and score for ordered rubrics, with probabilities and confidence or weighted scores. A Workers AI request can contain up to 64 questions and four images. The models also read text, JSON, images, and video within a 64K-token context, making them suited to multimodal classification and workflow automation.
Clef is post-trained from Qwen3.8-27B, while Clef-flash uses Qwen3.5-9B, and both retain a vision encoder. Inference uses a two-stage design: the backbone makes one prefill-only pass, then a small joint schema head reads the final hidden states. The head routes evidence to questions, lets fields cross-attend, and scores all options jointly before a per-question softmax produces probabilities. The goal is schema-aware, reproducible output without improvised text.
Cloudflare froze both backbones and trained the routing head with rank-256 low-rank adapters. Its loss combines label-smoothed cross-entropy with Brier loss to reward correctness and calibration, while Reinforcement Learning for Calibrated Decisions gives partial credit for adjacent ordinal choices. That emphasis could help automated routing, risk triage, and other decisions where uncertainty matters. It does not, however, remove the need to design reliable schemas and evaluate domain-specific errors.
On Cloudflareās Decision Index 0.2.1 shortlist, a Clef model ranked first on seven of ten benchmarks. Reported examples include 94.20 macro-F1 on BANKING77 versus Jevās 79.74, 97.43 on CLINC150+OOS versus 89.27, and Clef-flashās 97.73 on home appliances versus Jevās 52.27. Jev still leads on GPQA Diamond, MMLU-Pro, and BBH, indicating that structured decision ability does not automatically equal stronger general knowledge. Clef also won three of four TypeSafe workflow evaluations, but every number is vendor-reported and awaits independent replication.
Clef and Clef-flash are available through the Workers AI binding, REST API, and AI Gateway; their weights are on Hugging Face for self-hosting. Cloudflare reports median latency of 209.3 milliseconds for Clef, 38.8 milliseconds for Clef-flash, and 524.1 milliseconds for Jev. In a threat-intelligence test, Clef classified a domain in 2.2 seconds, compared with 4.7 seconds for gpt-oss-120b. These results suggest a useful speed advantage, though self-hosting was tested on a single H200 with BF16 weights and hardware needs will vary.
The available evidence invites a cautious reaction: Clef looks strongest on bounded classification and workflow tasks, while Jevās knowledge-benchmark leads remain significant. Cloudflare is also announcing a private-data reinforcement-learning service, beginning with forward-deployed engineers and later expanding into self-serve access through AI Gateway, Workers AI, Containers, and a new Trainer component. If the API proves stable, Clef could become an attractive backend for agents that need machine-readable decisions rather than chat. The outlook depends on independent validation, calibration across domains, and tooling that makes schema authoring and fine-tuning easier.
