How to get Clef: trial, API key, Workers AI and open weights

The four ways to run Clef

Clef is a decision model: it answers typed questions — Choice, Score or Noul — about a state you send, with probabilities rather than text. It comes in two sizes, Clef (27B) and Clef-flash (9B), and you can reach them four ways:

1. No-signup trial — Clef-flash in your browser, on preset use cases.
2. JevStation API key — Clef, Clef-flash and Jev behind one key, billed in credits.
3. Cloudflare Workers AI — your own Cloudflare account, billed per token, with a daily free allowance.
4. Open weights — download from Hugging Face and run it on your own GPU.

1. No-signup trial

Steps. Open the Clef playground, pick an example, edit the text and press Run. The trial above this section works the same way, and the main playground has the same model picker.

Price. Free: 3 runs a day per visitor, on the preset question sets. The larger Clef model is not part of the trial.

Suits a first look at what a decision model returns.

2. JevStation API key

Steps. Create an account, make an API key in Settings › API Keys, and call POST https://jevstation.com/api/v1/systemone with "model": "clef-flash" or "model": "clef". Without the field, the same call runs Jev.

Price. Credits per evaluation, not per token. Clef-flash costs the same as Jev: 1 credit for a standard evaluation (up to 8,000 characters of state and five questions), 3 for a larger one. Clef costs twice that, 2 and 6, because its token price is nearly six times Jev’s. New accounts get 200 free credits; packs are $9.90 for 5,000 and $15.90 for 12,000, and a failed evaluation is refunded.

Pros. Jev and both Clef models with one key and one request format, so you can compare them side by side on your own data; a playground for designing questions, batch runs over a CSV, and MCP tools for Claude Code and Cursor.

Cons. More expensive per call than Workers AI at list price, and there is no free daily allowance beyond the trial.

Suits teams that want to test Clef against Jev, or that do not deploy on Cloudflare.

3. Cloudflare Workers AI, directly

Steps. Create a Cloudflare account. From a Worker, add an AI binding and call env.AI.run("@cf/cloudflare/clef-flash", body). From anywhere else, create an API token with the “Workers AI” template (or Workers AI read and edit permissions) and call the REST endpoint with your account ID.

Price. $0.24 per million input tokens for Clef and $0.09 for Clef-flash, output free. The first 10,000 Neurons each day are free on every account — about 1,200 typical Clef evaluations or 3,300 Clef-flash evaluations, by our token measurements. Beyond that you need the Workers Paid plan.

Pros. The cheapest hosted Clef, especially at small volumes, and the lowest latency if your code already runs on Cloudflare. Cloudflare says it does not store or train on requests.

Cons. You write and maintain the integration, the REST response is wrapped in an envelope you have to unwrap, and there is no Jev on Workers AI.

Suits developers who already deploy on Cloudflare and only need Clef.

4. Open weights

Steps. Download Cloudflare/clef or the Clef-flash weights from Hugging Face. Clef is built on Qwen3.8-27B and Clef-flash on Qwen3.5-9B; both are licensed Apache 2.0.

Price. Your hardware. Plan on a data-centre GPU for the 27B model.

Pros. Data never leaves your machines, there is no per-call bill, and you control the version — and, unlike the hosted model, the context limit.

Cons. You run the inference stack, and Cloudflare’s reinforcement-learning fine-tuning service is not part of the download.

Suits regulated data and very high, steady volume.

Call Clef through JevStation

The request is the same one the Jev guide uses, plus one field. Replace $JEVSTATION_API_KEY with your key.

curl -X POST https://jevstation.com/api/v1/systemone \
  -H "Authorization: Bearer $JEVSTATION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "clef-flash",
    "state": "I have been trying to connect my Stripe account for 3 days and it keeps failing with invalid redirect. I am losing sales every hour.",
    "questions": {
      "category": {
        "type": "choice",
        "instructions": "Which team should handle this support ticket?",
        "criteria": {
          "billing": "Payments, invoices, refunds or plan changes",
          "integrations": "Connecting or syncing a third-party service",
          "bug": "The product behaves incorrectly",
          "account": "Login, access or account settings"
        }
      },
      "churn_risk": {
        "type": "noul",
        "instructions": "Is the customer at risk of churning?"
      }
    }
  }'

The response has the same shape whichever model answers: data.model, data.answers, data.usage, and data.credits.charged. GET /api/v1/systemone lists the models you can pass and what each costs. To run the same check from an agent, the MCP tools take the same optional model argument.

Call Clef on Workers AI directly

From a Worker with an AI binding:

const result = await env.AI.run('@cf/cloudflare/clef-flash', {
  model: 'clef-flash',
  state: ticketText,
  questions: {
    churn_risk: {
      type: 'noul',
      instructions: 'Is the customer at risk of churning?',
    },
  },
});
// result.answers.churn_risk.noul → 0 to 1

Over REST, the same body goes to https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/ai/run/@cf/cloudflare/clef-flash with Authorization: Bearer $CLOUDFLARE_API_TOKEN, and the answers come back under result. A model value that does not match the URL is rejected with error 5006.

Before you build on it

• Only the start of the state is read. On 2026-10-04 Workers AI read about the first 2,048 tokens of a Clef state and silently ignored the rest. Put the decisive text first. The full test.
• Retry once on 5xx. We saw one “Clef inference failed” (code 5012) in about 35 Clef calls.
• Benchmarks are the vendor’s. Cloudflare reports Clef ahead of Jev overall; no one has reproduced it yet. Jev vs Clef lists where each one leads.

Comparison at a glance

JevStation is independent and not affiliated with Cloudflare or TypeSafe. Clef is Cloudflare’s model; we run it through Workers AI.