Four answer types.
Yes/no, pick one, rate a result, or select all that apply. Ready to use in your application.
THE SELF-HOSTED DECISION MODEL v0.3
Intelligence,
decided.
Turn context into decisions.
A small AI model that reads once, answers many,
and runs on infrastructure you control.
Runs on one GPUNo text generationJev-compatible API
ONE TEXT. MANY ANSWERS.01 — 03
YOUR TEXTREAD ONCE
“I was charged twice for my subscription. Please refund the duplicate payment.”
ask your questions
Q01 yes / no
Needs a refund?
yes
Q02 pick one
Which team?
billing
Q03 pick many
Which topics?
payments, refund
Answers your code can useNo generated text
Example decisions · not a live model response
TEXT DECISIONS96.1%Matches expected answers in our tests AI RESPONSE REVIEW92.5%Checks quality, accuracy & safety READ YOUR TEXT ONCE1 readAcross every question in a request GENERATED TEXT0Answers your code can use
Accuracy on our project’s test questions, not a guarantee for every use case. What did we test? ↗
WATCH / ONE RECORDED RUN
One command to a live API.
Recorded end to end.
A real run on one NVIDIA L40S: deploy, six answers from one read of the text, Jev’s API, clear errors, and two fine-tuning jobs over HTTP. 14 of 14 checks passed.
01 / BUILT TO DECIDE
Some things need an answer.
Not another conversation.
Route a request. Check an agent’s work. Apply a policy. SelfJev turns your text, images, and questions into structured decisions with probabilities—without waiting for a generated response.
Your data stays yours.
Run the model on your own GPU. Process sensitive inputs without sending it to an external model API.
A familiar interface.
Already using Jev? Keep the same request format and point your application to your own server.
02 / HOW IT WORKS
One context.
Every angle.
A support message can raise several questions: what happened, who should handle it, and what to do next. SelfJev reads the message once and reuses that work for every answer.
01 ONE SHARED INPUT
Customer message
I was charged twice for my subscription. Please refund the duplicate payment.
Read once. Available to every question.
FOLLOW THE DECISION
“Please refund” gives the first question a direct answer.
02 ASK FROM EVERY ANGLE
FOLLOW THE DECISION
“Please refund” gives the first question a direct answer.
03 READY FOR YOUR CODE
refundtrueteam"billing"topics["payments", "refund"]01. Give it contextA message, document, or AI response, together with the questions you need answered.
02. Ask several questionsEach question uses the same text, so the model avoids reading it from scratch each time.
03. Act on the answersGet clear choices and probabilities to route, filter, or review in your own application.
DEFAULT / VISION + TEXT
One photo.
More answers.
The default SelfJev-4B is now trained on images and text. Ask about a photo and get choices and probabilities through the same API. Image input is a SelfJev extension: Jev accepts text, not images.
Python / Images
from pathlib import Path
from selfjev import SelfJev, Choice
# Uses your SELFJEV_API_KEY.
client = SelfJev(
base_url="http://localhost:8000",
)
result = client.system_one(
state=[Path("cat.jpg")],
questions={
"breed": Choice(
"What breed is it?",
{
"persian": "a Persian cat",
"siamese": "a Siamese cat",
"other": "another breed",
},
),
},
)Pass a Path, image bytes, or a PIL image. Over HTTP, use a base64 data:image/… URL as state, or a list of text and image parts.
SAME TREE. A VISUAL ROOT.See it once.
See it once.
Ask away.
Qwen3.5’s own frozen vision encoder processes the image once at the root. Every question and option branches from that shared work.
- Text · 1 question
- 136 ms
- Image · 1 question
- 163 ms
- Image · 5 questions
- 167 ms
Five questions, almost the same wait.
Median warm request times on one NVIDIA L40S, measured over 20 photos. The one-question image request is yes/no; five questions mix answer types. Model loading and network time are excluded. Measured before the vision fine-tune; your images and hardware will affect latency.
Read the timing recordVISION RELEASE / IMAGE EVALUATION
Teach it your world.
The default model already learns from images. Adapt it to your own photos with selfjev finetune or /v1/fine_tuning/jobs. Training rows accept images as state, too.
All image questions90.4%2,002 held-out questionsText-only baseline: 75.5%Familiar image kinds94.2%6 kinds used in training · new photosText-only baseline: 70.4%Unseen image kinds84.3%4 kinds excluded from trainingText-only baseline: 83.5%
Fine-tuning improved familiar image kinds, but showed no statistically significant transfer to unseen kinds. These are results on held-out photos, not a promise for every image task.
Trained kinds: pets, clothing, bean plant health, rice, satellite land use, and waste. Unseen kinds: snacks, indoor scenes, painting styles, and hurricane damage.
Text-only baseline report Find image training dataYOUR MODEL / YOUR REQUIREMENTS
Make room.
Make it yours.
Your documents, your vocabulary, your edge cases. Control the context budget and adapt the model to the decisions that matter to you.
A bigger foundation.
A limit you control.
Jev allows 32K tokens for your text plus the longest question. SelfJev has a configurable limit, built on a model with a 262,144-token native context window.
Jev · text + longest question32K
SelfJev · base-model capacity262K
SelfJev defaults to 32,768 tokens; larger windows need sufficient memory and validation. The full base-model window has not been validated in our engine. Jev also allows 64K across a whole request.
Context & sizingYour use case.
Your fine-tune.
Not getting the decisions you need? Fine-tune SelfJev on examples from your own workflow: your labels, your policies, your definition of a good answer.
- 01Show it what good looks like.Pair real inputs with verified answers.
- 02Train a lightweight adapter.Use the included supervised fine-tuning tools.
- 03Evaluate. Then deploy.Check held-out examples and serve your model.
03 / THE EVIDENCE
Strong results.
Receipts included.
Can it choose the right answer from a piece of text? We tested yes/no decisions, choosing between options, and selecting every answer that applies.
How often did the model choose the expected answer? The same 1,991 questions test yes/no decisions, choosing one answer, and selecting all that apply.
MODELANSWERS MATCHED
Scores are the share of answers matching the expected result; for “select all,” every choice must match. Only models with a recorded result are shown. Click a row for its report.
SelfJev approaches Jev’s accuracy on these tests and runs on your own infrastructure. The questions and expected answers were created and checked by AI; results on your own data may differ.
How we tested it↳Read the findings
The breakthrough was learning the right task.
Handling exceptions, weighing alternatives, and checking AI responses all needed targeted training. Improving those examples mattered more than simply making the model bigger.
04 / THE SPEED STORY
Less repeated work.
More decisions.
From a local Mac to a dedicated GPU. Explore measured response times, adjust your workload, and see what fits your infrastructure.
404ms
How we measured speed H100 · measured processing time
50 questions about one 2K-token text
AVAILABLE ON HUGGING FACE
Jwuthrich Get the weights. Explore the evidence.
DEFAULT / VISION + TEXT
SelfJev-4B
The image-and-text adapter used by the native tree engine. Downloaded automatically by SelfJev.
Get the model ↗MERGED WEIGHTSReady for vLLM
The complete vision release with tokenizer and configuration. vLLM serves text; images need the native engine.
Get merged weights ↗EVALUATION DATASETDecision Bench
3,657 questions across text decisions, AI response review, and record reasoning.
Explore the dataset ↗Looking for the previous model? Text-only release
05 / YOUR INFRASTRUCTURE
From a question
to your first call.
A lightweight Python SDK. One GPU to serve. Docker and an AWS deployment command, with practical guides for Runpod and Google Cloud.
Python / your first decision
from selfjev import SelfJev, Noul, Choice
# Use the secret you set as SELFJEV_API_KEYS on your server.
client = SelfJev(
base_url="http://localhost:8000",
api_key="your-server-key",
)
result = client.system_one(
state="I was charged twice. Please refund me.",
questions={
"refund": Noul("Does the customer want a refund?"),
"team": Choice("Which team should handle this?", {
"billing": "payments and refunds",
"support": "technical support",
}),
},
)