The Pulse
Cloudflare Releases Clef Decision Models and RL Fine-Tuning Service
Cloudflare has released Clef and Clef-flash, open-weight models that return typed decisions and probabilities rather than free-form text. The company is also offering customer fine-tuning with engineers before a planned self-service reinfor
AI.info Team · Written with AI ·
Cloudflare is pitching Clef as a faster alternative to general-purpose language models for decisions such as routing a support ticket, but the company’s own tests show a narrower trade-off: Jev, the decision model it challenges, still scores higher on one business workflow. Announced on October 1, Clef and the smaller Clef-flash return probabilities for predefined choices instead of generating free-form text. Cloudflare is releasing their weights under the Apache 2.0 license and pairing them with a new service to fine-tune Clef for customer workloads.
Clef trades open-ended answers for bounded choices
A decision model takes a state—such as a customer message—and a set of typed questions, then scores the permitted answers. Code can use those scores to route a request, escalate it or pass it to a person. Unlike a language model asked to produce a structured response, Clef does not first generate text that software must interpret.
Cloudflare says Clef uses Qwen as its base and performs a prefill pass before scoring valid schema choices in parallel. Its architecture combines a routing head with rank-256 low-rank adapters; the company says its training used label-smoothed cross-entropy and Brier loss to improve probability calibration. Clef also has a vision encoder, according to Cloudflare, and a 64,000-token context window—twice the 32,000-token window the company attributes to Jev.
Cloudflare reports a latency lead, with a workflow exception
In the company’s 43-evaluation latency comparison, Clef-flash recorded a median of 38.8 milliseconds and Clef 209.3 milliseconds, against 524.1 milliseconds for Jev. Cloudflare’s reported 95th-percentile figures were 122.4 milliseconds for Clef-flash, 238.6 for Clef and 536.0 for Jev. The results come from Cloudflare’s own evaluation run, not an independent audit.
On four business workflows in Typesafe’s evaluation suite, Clef scored above Jev on invoice processing, customer service and security incidents. Jev scored 71.6 on agent-trace observability, ahead of Clef at 68.5 and Clef-flash at 69.8. That split matters: the speed figures do not establish that either Clef model is the best choice for every decision task.
“When you fine-tune a model, you may give up some general purpose performance in exchange for higher accuracy in a specific domain.”
Michelle Chen, Alex Reneau and Kevin Flansburg, Cloudflare Blog authors
RL tuning starts with Cloudflare engineers
Cloudflare’s new reinforcement-learning offer is not yet a self-service product. Customers can work with the company’s forward-deployed engineering team to tune Clef for their workload; Cloudflare says it will use those engagements to build a self-service platform for collecting data, fine-tuning models and redeploying them on its infrastructure.
The proposed workflow joins several Cloudflare services: AI Gateway can capture request-and-response data, Workers AI can generate model rollouts, Containers can provide a sandbox to score and replay actions, and a new Trainer component updates the model weights. The company also plans redeployment through Workers AI. Cloudflare says internal teams are exploring uses in Trust & Safety review, support triage and bot classification.
Open weights do not remove the specialization trade-off
Developers can run the released weights locally or use the hosted models on Workers AI, and the models are compatible with Jev’s API. That compatibility could ease testing for teams already using Jev, while the fixed-choice output may suit workflows that need predictable fields and confidence scores.
The compromise is built into the product: Clef is designed to decide among options supplied by the developer, not to replace a model that must reason and respond freely. Cloudflare’s plan to start with engineer-led tuning reflects the work needed to shape a model around each customer’s labels and decision rules. For now, its self-service RL platform remains a future step, while the downloadable weights and hosted Clef models are available to try.