Clef Flash
Hosted on our own DGX Spark in Austria. Decision model: answers typed questions with probabilities, through its own endpoint instead of chat.
Try Clef Flash 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-flash:9b
- Size
- 9B dense (Q8_0)
- Context
- 256K
- Input
- Text, Image
- Capabilities
- DecisionImage
- Tier
- Decision
- Hosting
- always loaded
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-flash",
"state": "…",
"questions": { … }
}
What Clef Flash is good for
Real-time routing
Assign every incoming message to a team, a model or a workflow right away. Clef Flash is permanently loaded and answers without warm-up.
Moderation pre-filter
Clef Flash detects hate speech and insults almost as well as the larger Clef. Use it as a fast first filter and pass borderline cases to Clef or a person.
Many small decisions
Yes/no, choice or scale, thousands of times a day: only the input is billed, the answer comes as a number.
Classifying images
Sort receipts, photos or screenshots by type and content without parsing text.
What it is not for
- Clef Flash does not write text. For answers, use a chat model.
- Clef Flash is weak at prompt injection (F1 0.24). If attacks on your bot must be detected, use Clef.
- Clef Flash raises more false alarms than Clef. Where a false alarm is costly, such as automatic blocking, the larger model pays off.
Example: route a support ticket
One choice and one scale in a single call. The scale comes back as a number between the levels, plus the distribution.
curl https://allsmartllm.com/v1/systemone \
-H "Authorization: Bearer sk-allsmartllm-…" \
-H "Content-Type: application/json" \
-d '{
"model": "clef-flash",
"state": "Support ticket: Since this morning the whole office has no internet.",
"questions": {
"team": {"type": "choice", "instructions": "Who is responsible?",
"criteria": {"tech": "Technical", "billing": "Billing", "contract": "Contract"}},
"urgency": {"type": "score", "instructions": "How urgent?",
"criteria": ["can wait", "today", "immediately"]}
}
}'
Our measurements
We measure Clef Flash regularly on our own unpublished test sets and on public benchmarks, with the same questions as the larger Clef.
| Clef Flash | Clef | |
|---|---|---|
| Found | 92.1% | 92.1% |
| Precision | 90.3% | 98.6% |
| Clef Flash | Clef | |
|---|---|---|
| Inappropriate | 0.87 | 0.85 |
| Hate against groups | 0.80 | 0.87 |
| Clef Flash | Clef | |
|---|---|---|
| Inappropriate | 0.64 | 0.70 |
| Clef Flash | Clef | |
|---|---|---|
| Rule bypass | 0.24 | 0.65 |
For hate speech Clef Flash is on par with the larger Clef; it finds sensitive data just as often but raises more false alarms. It hardly detects prompt injection. It is the cheaper choice for fast pre-decisions, not for protection against attacks.
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 Flash run?
On our own DGX Spark machines in Austria, permanently loaded. Your requests do not leave the EU and are not used for training.
Clef or Clef Flash?
Clef Flash when speed matters and a false alarm is cheap. Clef when accuracy matters or prompt injection must be detected.
Can I call Clef Flash via /v1/chat/completions?
No. Decision models have their own endpoint: POST /v1/systemone.
Does Clef Flash understand images?
Yes. Images are sent as base64 (PNG, JPEG, WebP); image URLs are not supported.
What does a call cost?
Only input tokens are billed; output is always 0.