Clef
Hosted on our own DGX Spark in Austria. Decision model: answers typed questions with probabilities, through its own endpoint instead of chat.
Try Clef for free
Every new account starts with 10M free tokens, usable for this and every other model. No card, no subscription.
Fact sheet
- Model name in the API
- clef:27b
- Size
- 27B dense (Q4_K_M)
- Context
- 256K
- Input
- Text, Image
- Capabilities
- DecisionImage
- Tier
- Decision
- Hosting
- loaded on demand
Call
Decision model: its own endpoint instead of chat. You send a state and typed questions and get probabilities per question back.
POST /v1/systemone
{
"model": "clef",
"state": "…",
"questions": { … }
}
What Clef is good for
Routing and triage
Sort requests by intent, department or urgency before a chat model or a person answers. One request, several questions, one call.
Moderation and safety
Detect whether a message insults, incites hatred, tries to bypass rules or hints at a crisis, with a probability instead of a gut feeling.
Checking texts and images
Does a document contain sensitive data? Does a photo match the damage claim? Images go in directly (base64).
Forms and workflows
Yes/no, choice and scale questions whose answer goes straight into your code, without parsing a sentence.
What it is not for
- Clef does not write text. For answers, summaries or translations, use a chat model.
- Clef is not permanently loaded: the first call after a pause takes a few seconds. For many small, time-critical decisions, Clef Flash is the better choice.
- The answer is only as good as the question. Vague questions ("Is this problematic?") give vague probabilities; concrete questions with clear boundaries work best.
Example: classify a customer message
Three questions in one call: one choice and two yes/no questions. You get a probability per question; only the input is billed.
curl https://allsmartllm.com/v1/systemone \
-H "Authorization: Bearer sk-allsmartllm-…" \
-H "Content-Type: application/json" \
-d '{
"model": "clef",
"state": "Online shop bot. User: You are all useless, I want my money back!",
"questions": {
"request": {"type": "choice", "instructions": "What does the user want?",
"criteria": {"refund": "Money back", "info": "Information", "other": "Other"}},
"insulting": {"type": "noul", "instructions": "Does the user insult a person?"},
"escalate": {"type": "noul", "instructions": "Should a human take over?"}
}
}'
Our measurements
We measure Clef regularly on our own unpublished test sets and on public benchmarks. All values use the same questions for Clef and Clef Flash.
| Clef | Clef Flash | |
|---|---|---|
| Found | 92.1% | 92.1% |
| Precision | 98.6% | 90.3% |
| Clef | Clef Flash | |
|---|---|---|
| Inappropriate | 0.85 | 0.87 |
| Hate against groups | 0.87 | 0.80 |
| Clef | Clef Flash | |
|---|---|---|
| Inappropriate | 0.70 | 0.64 |
| Clef | Clef Flash | |
|---|---|---|
| Rule bypass | 0.65 | 0.24 |
Clef and Clef Flash find the same number of sensitive texts, but Clef raises far fewer false alarms. The gap is largest for prompt injection: Clef catches almost three times as many attacks. Clef Flash is on par for hate speech and cheaper to run at high volume.
F1: balance of recall and precision, 1 = perfect. Public benchmarks: Multilingual HateCheck (Röttger et al., CC BY 4.0), GermEval 2018 (Wiegand et al., CC BY 4.0), deepset/prompt-injections (Apache 2.0). Measured in November 2026 on our hardware.
Frequently asked questions
Where does Clef run?
On our own DGX Spark machines in Austria. Your requests do not leave the EU and are not used for training.
How does it differ from Clef Flash?
Clef is the larger model (27 billion parameters) and more accurate, especially regarding false alarms and prompt injection. Clef Flash (9 billion) is permanently loaded and answers without warm-up.
Can I call Clef via /v1/chat/completions?
No. Clef does not generate text and has its own endpoint: POST /v1/systemone. Chat requests get a 400.
How many questions fit into one call?
Up to 64. All are answered in a single pass; the time grows only slowly with the number of questions.
What does a call cost?
Only input tokens are billed; output is always 0. A typical decision with a short state is a few hundred tokens.