Simple Jev

What it does

Simple Jev, from Featherless AI, turns open Hugging Face language models into a Jev-style classifier. You send shared context and typed questions (Choice, Score, or Noul), and it reads the model’s next-token scores for each answer label to build probabilities, choices, and rubric scores, without generating text and without training a separate classifier head. /v1/systemone is an alias of its own /v1/classifier route.

The README says the API takes inspiration from TypeSafe’s interface “and includes project-specific behavior”, and it points to its own API reference as the contract. There’s also a public demo API serving larger models with no login, limited to 2k tokens of context and rate-limited. The repository includes RFDT, scripts for fine-tuning a model on one classification task, and a playable adaptation of JevPilot that runs on the demo API.

Setup

python -m pip install -e './hf-server'
python hf-server/hf_server.py --model Qwen/Qwen3.5-0.8B \
  --device cpu --dtype float32 --max-model-len 4096 --max-batch-size 4 --max-batch-tokens 4096

It needs Python 3.12 or newer. The model downloads on first run. For a GPU, install the right PyTorch build first. The README’s GPU example uses Gemma 4 26B-A4B.

Using it

We served Qwen3.5-0.8B on a four-core CPU with no GPU, and ran the same compatibility checks with TypeSafe’s official Python SDK (0.7.0) that we use for every Jev reproduction.

• The tests pass. All 22 server tests and 24 shared-module tests passed after installing the test extra.
• It runs on a CPU. The small Qwen model served every request on our machine.
• Core requests work through the SDK. All three question types, chat-message and JSON state, SDK response_model parsing, Noul criteria, and Choice options without descriptions all worked. The model field in responses names the model that answered.
• Several differences from Jev’s API:
  • The model field must be the exact Hugging Face ID the server was started with, so both jev-latest and pinned names like jev-1.13.0 are rejected. Code written for Jev needs its model name changed.
  • Choice and Score accept at most 50 options, where Jev’s Choice accepts up to 255. Our 64-option test was rejected.
  • instructions is required on every question, although it’s optional in Jev’s schema.
  • There’s no GET /v1/models, so the SDK’s models.list() fails.
  • There’s no API key check.

Overall it passed 6 of our 13 checks, including the extra alias test. That’s the lowest of the servers we’ve run this suite against, largely by design: it defines its own contract rather than cloning Jev’s.

Permissions and data

• Self-hosted, your data stays local. The server binds to 127.0.0.1 by default. It has no authentication, so don’t expose it without adding your own.
• The public demo is someone else’s server. Requests to the free demo API, and from the bundled driving demo, go to Featherless with no authentication. Don’t send private data to it.
• No license. The repository has no license file. The driving demo is an attributed adaptation of JevPilot, which has no license either.

Maintenance

Nineteen commits on September 18, with a documentation site, an API reference, tests, and the training scripts. One issue is open. It’s a young project from a company that sells model hosting, and the README points to paid plans for production use.

Verdict

Simple Jev is an easy way to try Jev-style questions on open models: the hosted demo needs no account, and the server runs a small model on a plain CPU. It works well through TypeSafe’s SDK for everyday requests, but it isn’t a drop-in replacement: model names, option limits, and the missing model list and authentication all differ. Treat it as its own API inspired by Jev, check the license question before building on it, and evaluate any model on your own labeled data.

For servers closer to Jev’s API, see jeff and openjev-sglang. For trained Jev-style models, see kev. Our Jev alternatives guide compares the approaches.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.