Open-weight 27B decision model that returns typed answers with probabilities for routing, triage and guardrails, not generated text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Where Clef Beats Jev on Classification and Where It Doesn't
Benchmarks
Clef on the Decision Index: Strong Routing, Weaker Reasoning
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
Clef or Clef-flash: Is the Precision Tier Worth It?
Clef and Clef-flash launched together on Workers AI. Both use the same System One request format and the same noul, choice and score question types. Cloudflare positions Clef as the precision model and Clef-flash as the hot-path model. Cloudflare reports Clef at 61.2 on its Decision Index against 57.1 for Clef-flash. Clef-flash is faster, at a median 13x faster than Jev against 2.5x for Clef, and has a lower token price. The practical decision is accuracy against speed and cost.
Choose Clef when decisions involve multiple intents, subtle image details or costly misroutes. In one hands-on test, Clef caught a case that Clef-flash missed. Choose Clef-flash when you need the lowest latency and cost for simple, high-volume routing. Cloudflare cut its price after launch, and it is now cheaper than Jev. Start with Clef-flash, then move individual questions to Clef where your labeled data shows errors. Neither model generates text.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Support ticket triage
Clef can decide in one request whether a ticket is urgent, which team owns it and how severe it is. This is Cloudflare's own documented example. Typesafe's workflow evals show Clef beating Jev on customer service. A single call can return up to 64 typed answers, which your code uses for routing or escalation. Add an "other" option to the team question, and split multi-topic tickets into yes/no checks.
Pre-action agent guardrails
Clef suits fast checks before an agent acts, such as whether a request is risky or a tool call could cause harm. The output is a calibrated-style probability your code can threshold, not prose. Cloudflare suggests placing Clef in the agent's request path and handing off to an LLM to take action. In independent testing it flagged a social-engineering request. Set thresholds on your own labeled data.
Visual and web classification
Clef has a vision encoder, so it can classify screenshots, receipts and forms, which text-only decision models like Jev cannot. Cloudflare's Threat Intelligence team paired it with Browser Run to fetch, render and classify a domain in 2.2 seconds, against 4.7 seconds for gpt-oss-120b in the same workflow. It accepts up to four images per request, and each image is capped at 1,024 tokens.
Invoice and security-incident classification
Cloudflare reports that Clef beats Jev on Typesafe's workflow evals for invoice processing and security incidents, as well as customer service. Structured JSON state fits records such as invoices or alerts. Ask several typed questions about one record, such as status, amount thresholds or severity, and get all the probabilities in one pass. These results are vendor-reported, so audit them before automating.