simple-jev
Featherless AI server that turns compatible open Hugging Face models into a /v1/systemone classifier by reading next-token logits.
Overview
Simple Jev from Featherless AI lets you use compatible open models from Hugging Face for structured classification and scoring without training a separate classifier head. You send shared context and a set of questions; the server reads the model's next-token logits for each question and builds a JSON response of choices, rubric scores or truth judgments, rather than having the model generate JSON. Shared validation, versioned prompt instructions and scoring rules live in a plain Python common folder so other inference stacks can reuse them.
The server runs with Transformers and PyTorch, optionally with a Laya backend, and exposes /v1/classifier with /v1/systemone as an alias; Choice supports up to 255 options and Score up to 50 levels. A public demo API serving Gemma 26B needs no key but is limited to 2k tokens and 2 requests per second. An evaluation framework includes native JevBench scoring, text and vision suites, resumable execution and offline unit checks, plus a reproducible prompt-format search. The API takes inspiration from TypeSafe's interface with some project-specific behaviour, and model compatibility is constrained.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What color is Mia's bicycle? | red · blue |
| Score | Rubric score questions (up to 50 levels) | — |
| Yes / no | Truth/support judgments | — |
How it works
- Start the server with a compatible Hugging Face model
- POST state and questions to /v1/classifier or /v1/systemone
- Server reads next-token logits per question
- Return choices, scores or truth probabilities as JSON
Strengths
- No classifier head training
- Free public demo endpoint
- Evaluation tooling with JevBench scoring
Caveats
- Only some models are compatible
- Demo limited to 2k tokens and 2 RPS
- Behaviour differs from TypeSafe's API in places
Getting started
curl https://simple-jev-demo-api.featherless.ai/v1/models, or clone the repo and python -m pip install -e './hf-server' then simple-jev --model <hf-model>.At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- HTTP server (Transformers/PyTorch)
- Has code
- yes
- Has tests
- yes
- Stars
- 602
Use cases and integrations
How curated lists describe it
Serving: turns any open model into a Jev-compatible classifier endpoint, so a self-hosted model answers the same typed questions as the hosted API, with 27 tests.
yibie/awesome-jevFeatherless AI's server that turns compatible open Hugging Face models into a `/v1/systemone` classification endpoint without training a classifier head, with a Laya backend and a reproducible prompt-format search
AnotiaWang/awesome-decision-modelsAdapter turning open LLM endpoints into Jev-compatible classification services without training a separate classifier head.
logicrw/awesome-jev-projectsTurn any open model into a classifier/jev endpoint
wh000wh000/awesome-jev-liveTransforms compatible open-model logits into typed decisions without a separately trained classifier head; model compatibility is constrained.
cobanov/awesome-jevTurn any open model into a classifier/jev endpoint.
hellogumbo/awesome-jevTurns open Hugging Face models into a Jev-style classifier endpoint by reading next-token logits, with a public demo API.
MrJev/awesome-jev
Frequently asked questions
Do I need to train a classifier?
No; Simple Jev reads next-token logits from compatible open models.
Is there a free demo?
Yes, a public demo API with a 2k-token context and 2 RPS limit.
Is it compatible with Jev clients?
/v1/systemone is an alias, but the project's own API reference is the contract.
Related projects
Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGoGo gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGoTurns any chat model on SGLang, vLLM, transformers or MLX into a Jev-compatible service with one prefill and one label token per question.
ProjectbetaPythonSelf-hosted Jev-compatible System One server that sits in front of llama-server and reads option probabilities from any GGUF model.
ServicebetaPythonInference engine for Jev-style decision models using segmented forwards with prefix reuse, serving the Jev wire contract.
ProjectresearchPythonResearch preview: Python client and pinned adapters for running System One decision models on your own Modal GPUs.
LibraryresearchPython