Clef: Open-weight decision models, and new RL fine-tuning platform

TL;DR Highlight

Cloudflare releases Jev-compatible decision models Clef and Clef-flash, claiming the top Jev Decision Index spot

Who Should Read

Backend or agent developers who want a cheap classification step ahead of LLM calls for ticket routing or content filtering

Core Mechanics

A decision model returns classifications with probabilities
- You pass inputs and questions and get typed answers with probabilities.
- Your code uses them to route tickets, escalate, or defer to a human.
Clef and Clef-flash are Jev API compatible
- Both are hosted on Workers AI and open-weighted on Hugging Face under Apache 2.0.
- Clef is the precision model and Clef-flash is for latency-critical decisions.
Differences from Jev: images and 64k context
- Clef has a vision encoder and can classify images.
- Its context window is 64k versus Jev's 32k.
Benchmarks claimed ahead of Jev
- BFCL case exact is 98.47 for Clef versus 95.75 for Jev.
- Clef beat Jev in 3 of 4 areas on Typesafe's own eval suite.
- On When2Call accuracy, Jev leads 80.97 versus Clef's 72.37.

Community Reactions

Opinion that it costs more than Jev
- Commenters calculate $0.24 per million input tokens for Clef, about 6x Jev's $0.042.
- Clef-flash at $0.09 is seen as far more competitive.
Experience that it was slower than Jev
- One test showed clef-flash at 661ms median versus 230ms for Jev.
- A knowledge-base write review test showed hosted p50 of about 850ms for Clef versus about 110ms for Jev, with similar quality (recall 0.98 vs 1.00).
Experience that quality fell short too
- For hate speech filtering, one user found Clef 2-3x slower and catching less than Jev.
- On coin image classification, an 8-bit quantized Clef scored about 41% versus 53% for local gemma4:27b, which ran 4x faster.
Open weights, not open source
- One comment says data and training pipeline are not published.
- Another comment says Jev-style models are of little use without fine-tuning because priors can't be passed in.

How to Apply

Compare Clef latency against your current LLM classifier
- Pick one classification call, such as routing or moderation, and send the same 100 inputs to Clef, Clef-flash and your current model.
- Compare median and p95 latency and accuracy.
Put the cheap classifier first and the LLM second
- Handle the request directly when the probability is above a threshold and defer to an LLM or a human otherwise.
- Log the fallback rate and error rate to tune the threshold.
Calculate cost with your own traffic
- Multiply input tokens per call by the price to get the monthly cost.
- If traffic is high, compare against self-hosting the Hugging Face weights.

Try It

copy & run

# Paste into a shell (based on the article's curl example; fill in your account ID and token)

export CLOUDFLARE_ACCOUNT_ID={{account_id}}
export CLOUDFLARE_AUTH_TOKEN={{token}}

curl https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/cloudflare/clef \
  -X POST \
  -H "Authorization: Bearer $CLOUDFLARE_AUTH_TOKEN" \
  -d '{"model":"clef","state":"Checkout has been failing for every customer for the last hour.","questions":{"urgent":{"type":"noul","instructions":"Is this support request urgent?"},"team":{"type":"choice","instructions":"Which team should handle this request?","criteria":{"billing":"Payments, invoices, and refunds","technical":"Outages, errors, and configuration","sales":"Plans and upgrades"}},"severity":{"type":"score","instructions":"How severe is the customer impact?","criteria":["No impact","Minor","Major","Critical"]}}}'

# Next: replace state with a real inquiry from your service and switch "model" to clef-flash to compare latency.

Terminology

decision modelA model that returns probabilities over fixed choices instead of free text. It's like a judge that scores the branches of an if statement in your code.

prefillThe step where the model reads the whole input at once. Skipping token-by-token answer generation makes it fast.

LoRA adapterA small trainable layer added while the big model stays frozen. Like sticking notes on a book rather than rewriting it.

Brier lossA loss function measuring how far predicted probabilities are from actual outcomes. It helps keep probabilities well calibrated.

open weightsThe model weights are released, but the training data and training code are not.