decider
Open family of System One-style models fine-tuned from Qwen3.5 and Gemma that return typed decision probabilities in one pass.
Overview
decider is a family of language models that never generate text. Given a state and typed questions, one forward pass returns a probability distribution for every question: Choice over 2 to 255 options, Score over 2 to 10 described levels, or Noul as the probability of yes. There is no decoding or parsing and no output outside the options you defined.
The smallest model, decider-2b, is a Qwen3.5-2B fine-tune installable with pip; larger variants include a 12B Gemma-based model and a 31B model served with vLLM, with the README citing JevBench and Decision Index scores for them. The project ships an HTTP server, GGUF loading, ten text games and a Super Mario Bros agent, plus recorded demos such as Tetris, where the harness shortlists eight placements and the model picks one, reported at 20 lines per game versus 0.6 for a random pick from the same eight.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team? | billing · technical · sales |
| Choice | Which of the shortlisted game moves should be played? | — |
How it works
- pip install decider-ai and load a checkpoint
- Call decide(state, questions)
- Receive choice, confidence and per-option probabilities
- Threshold or act on the result in code
Strengths
- Open weights in several sizes
- Runs on CUDA, MPS or CPU
- One forward pass per request
- Includes server and GGUF support
Caveats
- Benchmark figures are the author's runs
- Larger models need large GPUs
Getting started
pip install decider-ai, then Decider("Mapika/decider-2b").decide(state, questions).At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- library, HTTP server
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open models: reproduces the System One shape with a Qwen3.5-2B fine-tune that emits typed decisions with calibrated probabilities in one pass.
yibie/awesome-jevQwen3.5-2B fine-tune that emits typed decisions with calibrated probabilities in one pass
AnotiaWang/awesome-decision-modelsA family of System One-style models fine-tuned from Qwen3.5, designed for…
wh000wh000/awesome-jev-liveOne-pass typed decisions with calibrated probabilities (System One style model), fine-tuned from Qwen3.5-2B.
hellogumbo/awesome-jevA family of System One-style models that never generate text: one forward pass over a state and typed questions returns a probability distribution per question. Ships ten text games and a Super Mario Bros agent where each move is one typed decision over the legal actions.
MrJev/awesome-jev
Frequently asked questions
Is decider a Jev clone?
It reproduces the System One shape with open fine-tuned models; it is not TypeSafe's model.
Which model should I use?
decider-2b for local use; the README recommends 12B or 31B variants for harder decisions on bigger GPUs.
Does it need an API key?
No, weights download from Hugging Face and run locally.
Related projects
Apache-2.0 395M ModernBERT System One model serving calibrated Choice, Noul and Score locally as a drop-in /v1/systemone replacement.
ServicebetaPython, TypeScriptGemma 3 4B option scorer that prefills context once and scores all options in one pass, on MLX or PyTorch, with a Doom demo.
ServiceprototypePythonGemma 4 12B with its LM head replaced by a small answer head, serving Jev-style noul, choice and score decisions with custom code.
ModelresearchPythonIndependent MLX port of Laya for Apple Silicon: typed decisions in about 13 ms (7 ms multilingual) with no PyTorch or cloud API.
ServicebetaPythonDeploys DiffusionGemma-Jev behind a TypeSafe-compatible API on a Cloud Run GPU, with snake, dino and tetris demos.
ServiceprototypePythonMangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython