SelfJev

SelfJev turns text, images and questions into decisions your code can use. Route a request, check an AI response, or apply a policy. Get typed answers and probabilities back from selfjev-4b, an open 4B decisions model with Jev's API.

SelfJev is self-hosted. There is no SelfJev cloud: you run the model server on your own GPU, and this package is the client for it (plus the server itself, as an extra). Already on Jev? Keep TypeSafe's SDK and change two environment variables.

1. Run your server

On a Linux machine with an NVIDIA GPU (24 GB is a practical start; see the hardware guide):

pip install "selfjev[serve,gpu]"
export SELFJEV_API_KEYS="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"   # your key, you choose it
selfjev serve --host 0.0.0.0 --port 8000

The first start downloads the selfjev-4b vision adapter (230 MB; text and images) and its Qwen3.5-4B base from Hugging Face. --adapter <dir or repo> serves your own fine-tune. Docker, AWS, Runpod and GCP recipes are in the deployment guide.

2a. Already using Jev? Change two variables

Code written for TypeSafe's typesafe-sdk runs unchanged against your server:

export TYPESAFE_BASE_URL="https://your-selfjev-host:8000"
export TYPESAFE_API_KEY="the key you set in SELFJEV_API_KEYS"

from typesafe_sdk import Choice, Noul, TypeSafeClient

client = TypeSafeClient()   # reads the two variables above
res = client.system_one(
    state="I was charged twice. Please refund the duplicate payment.",
    questions={
        "refund": Noul(instructions="Does the customer want a refund?"),
        "team": Choice(instructions="Which team?", criteria={"billing": "payments and refunds", "support": "technical issues"}),
    },
)

The default model name jev-latest is answered by selfjev-4b, and client.models.list(), errors and retries behave as they do against Jev. OpenRouter's decisions path (/api/alpha/decisions) is served too.

2b. Or use the selfjev client

pip install selfjev in your application. It adds Multi (select all that apply) and the fine-tuning API.

from selfjev import Choice, Multi, Noul, SelfJev

client = SelfJev(base_url="https://your-selfjev-host:8000", api_key="the key you set in SELFJEV_API_KEYS")

result = client.system_one(
    state="I was charged twice. Please refund the duplicate payment.",
    questions={
        "refund": Noul("Does the customer want a refund?"),
        "team": Choice("Which team should handle this?", {"billing": "payments and refunds", "support": "technical issues"}),
        "topics": Multi(
            "Which topics are mentioned?",
            {"payment": "a payment or charge", "refund": "a refund request", "login": "an account access problem"},
        ),
    },
)

print(result.nouls["refund"].noul)   # probability of yes
print(result.choices["team"].choice)  # selected team
print(result.multis["topics"].multi)  # selected topics

AsyncSelfJev has the same interface with await. SELFJEV_BASE_URL and SELFJEV_API_KEY work in place of the arguments.

Images (new in 0.3.0; Jev takes none): put a photo, a screenshot or a scanned document in state, alone or next to text. A Path, image bytes or a PIL image all work; on the wire it is a base64 data URL.

from pathlib import Path

result = client.system_one(
    state=[Path("cat.jpg")],
    questions={"breed": Choice("What breed is it?", {"persian": "a Persian cat", "siamese": "a Siamese cat"})},
)

The image is read once and every question is answered against it: 163 ms for one image and one question on an L40S, 136 ms for text. Fine-tuning takes image rows too.

Measured

SelfJev-4B Vision (the default since 0.3.0) served by TreeServer, compared with Jev on the same questions:

These are project evaluations, not a universal ranking. See the results and limitations and the Decision Bench evaluation set.

Links

• API reference · Fine-tuning · Deployment
• Model: selfjev-4b vision · text-only · merged weights (text-only)
• Source · Research notebook

The package code is Apache-2.0. The model weights carry their own licenses; see each model card.