Clef 27B

Cloudflare/clefserverless

A 27B decisions model for realtime classification, routing, and scoring: up to 64 questions answered about one input in a single call, billed on input tokens only.

vision
Input
$0.24 / 1M
Context
64K
Type
Decisions
Maker
Cloudflare
Released
2026-09-30

Positioning

Clef is not a chat model. It reads one input (a support ticket, an email, a review, a transaction note) together with every question you need answered, and returns probabilities instead of generated text. Use it at the front of a pipeline where you would otherwise burn chat tokens on classification: routing, triage, prioritization, moderation, extraction of discrete fields. It answers up to 64 questions about the same input in one call, so a full routing-plus-urgency-plus-sentiment pass costs one request, not three.

In the sference portfolio it complements the chat models rather than competing with them: Clef does the cheap, deterministic front-of-pipeline work, and the chat models do the generation behind it. If you find yourself prompting a chat model to output only a single word or a JSON enum, that workload belongs here.

Upstream, Clef accepts image input as well. Serving images on sference's decisions endpoint is in the works — until it ships, the API accepts text-only inputs, and the catalog capability flag will flip when vision goes live.

Compared to the portfolio

ModelContextInputOutputIndexVision
GLM 5.31M$1.20$4.2045—
DeepSeek V4.1 Flash (1017)1M$0.35$1.3245✓
GLM 5.3 Flash1M$0.20$0.6042✓
DeepSeek V4 Flash (0731)1M$0.28$0.5634—
Clef 27Bdecisions64K$0.24——✓

Serverless rates per 1M tokens at the realtime (list) rate — priority bills 1.3×, flex 0.7×, and batch 0.5× of it, and requests that arrive off-peak bill at a discount (see off-peak pricing) — confirmed per model when your account is provisioned. Dedicated deployments are billed per GPU-hour on monthly commitments.

Use it

POST /v1/decisions
Realtime only; no batch, no streaming
curl https://api.sference.com/v1/decisions \
  -H "Authorization: Bearer $SFERENCE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Cloudflare/clef",
    "state": "Please refund the duplicate charge.",
    "questions": {
      "urgent": { "type": "noul", "instructions": "Does this require immediate action?" }
    }
  }'

Decisions guide · API reference

Request limits

LimitValue
Context65,536 tokens (state + encoded questions/schema)
Questions≤ 64 per request
Options≤ 256 per question
Request body13 MiB
Latency target2 s admission; 20 s hard deadline; 21 s API timeout
Prefix cacheOff
BillingSuccessful responses only; usage aggregated in 5 s flushes

Question types

  • choice — pick one criterion → label + probabilities per option
  • score — ordinal criteria → expected 0-based index
  • noul — yes/no → P(true) only
  • confidence — the largest per-question probability, returned alongside every answer

Example request

{
  "model": "Cloudflare/clef",
  "state": "Please refund the duplicate charge.",
  "questions": {
    "route":  { "type": "choice",
                "criteria": { "billing": "Payments and refunds",
                              "support": "Technical help" } },
    "urgent": { "type": "noul",
                "instructions": "Does this require immediate action?" }
  }
}

Serving stack

  • Weights: simonlehmann/clef-NVFP4 @ 817ac58a
  • Reference: Cloudflare/clef (Transformers BF16) @ 2f3de3dd
  • Precision: backbone NVFP4/FP8 mixed · head + option embeddings BF16 · softmax FP32

Resources

FAQ

When should I use Clef instead of a chat model?

Whenever the output you need is a label, a score, or a yes/no answer rather than generated prose. Clef is cheaper (input tokens only), faster (no autoregressive decode), and returns calibrated probabilities per option, which chat models do not. If your prompt asks a chat model to reply with a single word, that is a decisions workload.

How is Clef billed?

On input tokens only: $0.24 per 1M tokens, and only for successful responses. There is no output rate because decisions models generate no text; usage is aggregated in 5-second flushes.

Can Clef see images?

Upstream Clef accepts image input. On sference, serving images on the decisions endpoint is in the works; until it ships the endpoint accepts text-only inputs, and the model card's vision capability will update automatically when it goes live.

What happens on the hard deadline?

Requests admitted under the 2-second latency target are given a 20-second hard deadline; if a response is not ready by then the request fails with a timeout and is not billed. Retry with the same input — decisions are stateless.

What precision is Clef served at?

NVFP4, the format of Cloudflare's native released checkpoint. We serve the weights as published, with no extra quantization on our side.

Where is data sent to Clef 27B processed, and is it used for training?

Every request runs on European GPUs. Inputs are never used for training, contractually and technically, and are deleted per your configured retention policy; compliance logs (model version, region, timestamps) are kept for your audit trail. Details in the DPA and sub-processor list.

What are the rate limits for Clef 27B?

There is no general per-account request limit, and the API does not send rate-limit headers. Image input has a per-team budget of 20 images per second; exceeding it returns 429. See Images and vision.

What happens when Clef 27B is updated or retired?

Pin an exact version by calling its versioned id: that version is the one that serves, and every response records it. Before a model retires we email you with a date, and the old id becomes an alias of its successor so existing integrations keep working. Rerun your evals against the successor before the date.

Generated from the live catalog (GET /v1/models), so pricing, context, and capabilities always match the API. Machine-readable version: https://sference.com/models/Cloudflare/clef.md