system-one-open
Open Jev replica on Gemma 4 E2B and Gemma 3 270M served on Modal, with eight demos and a TypeSafe eval head-to-head.
Overview
System One, open is an MIT-licensed replica of TypeSafe's Jev built on Gemma 4 E2B (attention LoRA) and Gemma 3 270M, trained and served on Modal. Like Jev it takes state and typed questions and returns choice, score and noul probabilities in one forward pass with no decoding, and it can handle up to 52 options in a single pass by using single-token letters.
The README compares the replica with published Jev numbers on a rebuilt TypeSafe public eval, reporting 76.7% for the replica against 86.9% for Jev on a strict common subset, plus results on held-out task types and email classification. Eight demos recreate public showcases: TypeSafe preset scenarios, a launch-video race, an inbox classifier, a viral-post analyzer, Doom via ViZDoom, a browser-use agent on Google Flights, a self-driving sim and a smart-home command parser. It is for people who want open weights and a self-hosted alternative and want to see where such a replica falls short.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option best fits the state (up to 52 options per pass)? | — |
| Score | Where does the state fall on an ordered set of levels? | — |
| Yes / no | Is this statement true of the state? | — |
How it works
- Send Jev-style or schema-style JSON to POST /decide.
- The Gemma-based model scores all options in a single forward pass.
- Return distributions, confidence and expected levels.
- Demos map the answers to actions such as game moves, browser operations or device commands.
Strengths
- Open weights and self-hosting on your own GPU or CPU.
- Single forward pass, no decoding.
- Published head-to-head against Jev with caveats in a full report.
Caveats
- Lower accuracy than Jev on the reported TypeSafe eval subset.
- Hosted demo scales to zero, so the first request after idle is slow.
- Hugging Face weight upload was still pending per the README.
Getting started
Run the Modal scripts (for example modal run typesafe_eval.py) or call the hosted demo endpoint described in the README.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Modal (FastAPI)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Open replica of TypeSafe.
wh000wh000/awesome-jev-liveOpen replica of TypeSafe's Jev: typed calibrated decisions in one forward pass, on Gemma 4 E2B / Gemma 3 270M (Modal).
hellogumbo/awesome-jev
Frequently asked questions
How does it compare with Jev?
The README reports 76.7% versus Jev's 86.9% on a strict common subset of TypeSafe's public eval, with the full report linked.
What can it decide?
Generic choice, score and noul questions; the demos use it for email triage, game actions, browser steps, driving decisions and smart-home commands.
Does it need a TypeSafe key?
No, it runs its own open model on Modal or your hardware.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Reads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
ProjectresearchPythonOpen Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.
ProjectresearchPythonSmall experiment: reordering enum options changed a local Jev clone's answers on its old default, while hosted Jev's did not.
DemoresearchPython- JevK5★ 150
Apache-2.0 open-weight Jev alternative that answers yes/no, choice and score questions with per-option probabilities in one pass.
LibrarybetaPython Open System One decision model (LoRA on Qwen3.5-4B) that reads typed answers from logits in one pass, with its benchmark harness.
ModelbetaPython