Cloudflare's 9B decision model returns typed probabilities in tens of milliseconds for routing and agent guardrails, not text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Built for Latency: 38.8 ms Median, Weaker on Hard Cases
Technical Specifications
What the model supports
Comparison
Clef-flash or Clef: Which Decision Model Fits Your Router?
Clef and Clef-flash share the System One API, the typed question schema, image input and the Apache 2.0 license. The decision is size. Clef is the 27B precision model; Clef-flash is the 9B latency model. Cloudflare reports launch-time median latencies of 38.8 ms for Clef-flash and 209.3 ms for Clef, though hosted Clef has since been made up to 2x faster. Hosted Clef keeps a 64K context window, while hosted Clef-flash is limited to 24K tokens.
Choose Clef-flash when labels are few and stable, inputs are short, and latency or volume cost dominates. Cloudflare's customer service workflow scores for the two are within a point. Choose Clef when option lists are long or "none of the above" is common (CLINC150 with out-of-scope examples: 97.4 vs 66.8), when you need grounding checks (RAGTruth: 79.4 vs 35.6), or when state exceeds 24K tokens. Scores are Cloudflare-reported; test on your own data.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Agent hot-path routing
Cloudflare positions the Clef family for the request path of an agent: classify first, then hand off to an LLM for the action. A 38.8 ms reported median and one-pass scoring make Clef-flash a fit for choosing which agent, tool or model handles a request. The limit is the 24K hosted context, so pass summaries rather than full histories.
Support ticket triage
Urgency (noul), team (choice) and severity (score) map directly onto Clef-flash's question types. On Typesafe's customer service workflow eval it scores within a point of Clef, per Cloudflare. It fits stable label sets and short tickets. Add an "other" option, and move to Clef when intent lists run to dozens of categories.
Tool-call and content guardrails
Because the model returns calibrated-style probabilities rather than text, your code can set its own block, review or allow thresholds. Cloudflare cites tool-call moderation, asking whether a call could cause harm, answered in tens of milliseconds before the tool runs. Treat it as a first-pass filter. It is not a grounding or fact-checking layer.
Image-based tagging
Clef-flash keeps the Qwen vision encoder and accepts up to four embedded images per request, which text-only decision models cannot do. Receipts, screenshots and product photos can be tagged with typed questions. In a small independent test, the larger Clef missed less on fine image detail, so verify legibility-sensitive tasks before relying on Flash.