choice
Choose from options
Supply the possible teams, actions or categories. You get a choice and a probability distribution over the options.
Who handles this request?Billing · Technical · Other
HostYourAI / Decision models
Models like Jev,hosted in the EU.
Use open models to sort customer emails, organise documents and flag urgent requests. You set the possible answers. The model helps your software choose the right one.
Compatible with the official Jev/TypeSafe SDK for Choice, Noul and Score. Use your normal HostYourAI API key and credits.
See how it worksIllustrative example
01 / What comes in
I was charged twice for the same invoice. Could you look into it?
Which team should handle this?
Billing
Technical
Other
Route to: Billing
These are example numbers, not model output. Choose a message or move the threshold. A probability is no guarantee; validate thresholds on your own data.
◎ Processing in the EU{ } Zero data retention↗ Jev/TypeSafe SDK compatible
01 / What are these models?
You set the options. The model helps choose.
Take a busy support inbox. You do not have to write a rule for every phrase a customer might use. Give the model the email and your teams. It returns the likely destination and a score for each alternative.
noul
Assess a yes/no question
The answer is a number between 0 and 1: the estimated probability of yes. You set the threshold at which your software acts.
Is this customer asking for a refund?An estimated probability of yes
score
Rate against a scale
Describe the levels of your scale. The model returns probabilities per level and a weighted score, for example to assess a message’s urgency.
How urgent is this message?Can wait · Today · Immediately
How this differs from a chat model
A chat model can write a reply. A decision model helps decide which team should handle the email and how urgent it is. Your software gets the agreed answer types and numbers directly. The two kinds of model work well together.
What the probabilities tell you
A defined answer structure prevents invalid labels, but the model can still pick the wrong option. Include an “other” option and send ambiguous cases for review. Check whether the reported probabilities match observed error rates.
02 / Inside your application
From an incoming message to the right place.
01
Read the situation
A customer email, an invoice or a request through your website.
02
Ask focused questions
Which category fits? Is there enough information to continue?
03
Choose the next step
Your application puts the email in the right queue. If the result is uncertain, a colleague takes a look.
A decision model can work alongside a chat model. A language model often remains useful for writing an answer or working through a complex problem.
Security and data
EU hosting, with control over your data.
For teams with privacy and compliance requirements: processing on our EU workers, restricted access and clear data processing terms.
Your input stays in the EU
The decision route uses only the configured EU worker. If it is unavailable, the request fails. It does not automatically fall back to an AI service outside the EU.
Access through your account
Use an API key from your account. Existing model restrictions and usage budgets apply. Direct GPU access is restricted.
Zero data retention
Jev-compatible models, without retaining your content. Your inputs and model outputs are processed to answer your request and are not stored. We retain only usage and technical metadata for billing, security and operating the service.
03 / The models
Choose a model that fits the job.
From small, fast models to our own Beslis Loes. Compare their performance and try them through the same API.
JevBench-NL public-120. Our Dutch translation of 120 public JevBench questions: 48 easy and 72 original. The large percentages show how many questions each model answered correctly. All models received the same 120 questions, measured on September 21, 2026.
Hosted in the EU Ready now
Beslis Loes
98.3%118 / 120 correct · Dutch test
JevBench (EN): no external score
Hosted in the EU
Bespoke Nimble
95.8%115 / 120 correct · Dutch test
JevBench (EN): 61.8 / 100
Hosted in the EU
SemIf · Qwen3.5-4B
90.8%109 / 120 correct · Dutch test
JevBench (EN): 74.7 / 100
Hosted in the EU
Decider · 2B
90.8%109 / 120 correct · Dutch test
JevBench (EN): 64.6 / 100
Hosted in the EU
Reflex · 4B
89.2%107 / 120 correct · Dutch test
JevBench (EN): 71.7 / 100
Hosted in the EU
SemIf · Qwen3.5-2B
80.0%96 / 120 correct · Dutch test
JevBench (EN): no external score
Hosted in the EU
Jeff / GLiFormer
70.8%85 / 120 correct · Dutch test
JevBench (EN): 66.9 / 100
Hosted in the EU
Laya
51.7%62 / 120 correct · Dutch test
JevBench (EN): 70.1 / 100
Hosted in the EU
NanoJev
31.7%38 / 120 correct · Dutch test
JevBench (EN): no external score
The displayed score measures accuracy only and is not weighted by speed or cost.
New Beslis Loes
Our Dutch alternative to Jev.
Our own Loes, now for decisions too. Have her sort Dutch customer emails, assess documents or flag urgent requests. You supply the options; Loes returns the probabilities.
The same API, SDK and credits. You can also combine it with another model in an ensemble.
98.3%118 / 120 questions correctJevBench-NL public-120
Try Beslis Loes
04 / For developers
Compatible with the official Jev/TypeSafe SDK.
Use the official Python SDK for System One requests with Choice, Noul and Score. Change the base URL, API key and model name. We have tested these calls end to end with TypeSafe SDK 0.7.0 on our EU GPU, including a two-model ensemble.
The decision API has its own endpoint alongside the chat API. Check the model limits before switching: SemIf currently supports 16 questions, 2–16 choice options and 4,096 input tokens per question. A matching API does not mean models make identical decisions.
Combine models in an ensemble
Have two models assess the same message. We combine their probabilities into one response, so you can test whether the pair works better than either model alone. If a model fails, you get an error rather than a silently incomplete result.
request.jsonEU GPU · beta
POST /api/decision/v1/systemone
{
"model": "hyai/semif-qwen3.5-4b",
"state": "I was charged twice.",
"questions": {
"department": {
"type": "choice",
"instructions": "Which team handles this?",
"criteria": {
"billing": "Invoices and payments",
"technical": "Bugs and integrations",
"other": "Anything else"
}
},
"refund_requested": {
"type": "noul",
"instructions": "Is a refund requested?"
}
}
}View an illustrative response
{
"model": "hyai/semif-qwen3.5-4b",
"answers": {
"department": {
"type": "choice",
"choice": "billing",
"confidence": 0.88,
"probabilities": {
"billing": 0.92,
"technical": 0.03,
"other": 0.05
}
},
"refund_requested": {
"type": "noul",
"noul": 0.24
}
},
"usage": {"input_tokens": 64, "output_tokens": 0}
}05 / EU hosting
Start with your own use case.
Start with examples your team can judge. Compare the model’s choice with the right action, and decide when a person should step in. Test models in the Model Garden with your normal API key. Usage comes from the same credit balance as other inference. An ensemble adds the usage of both models.
Open the Model Garden ↗Will you host the original Jev model?
TypeSafe’s Jev weights are not publicly available for self-hosting. This page covers independent projects that perform similar decision tasks. HostYourAI is not affiliated with TypeSafe.
Is this a drop-in replacement for Jev?
The API is compatible with the official Jev/TypeSafe Python SDK for System One requests with Choice, Noul and Score. Change the base URL, key and model name. Check model limits and retest your application. Decisions, calibration and Score confidence are not identical to Jev.
Where will my data be processed?
Models run on our own GPU workers in the EU. The Model Garden lists its region next to the model. Input travels over an encrypted connection to that worker, without falling back to another AI provider.
Can I use Jev-compatible models with zero data retention?
Yes. Zero data retention applies by default to the inputs and model outputs of our decision API, including requests made with the official Jev/TypeSafe SDK. Processing runs on our own GPU workers in the EU. Only usage and technical metadata are retained, such as token counts, costs and request status.
Does this work in Dutch?
Yes. Beslis Loes is our own Dutch decision model. It answered 118 of the 120 Dutch questions in our test correctly (98.3%). All other hosted models were tested on the same questions too. Compare the Dutch scores ↗
Jev and TypeSafe are names of TypeSafe AI. The projects listed here are independent.