decider-2b
Open replica of Jev System One: a 1.9B Qwen3.5 model that returns option probabilities in one forward pass—no decoding, no JSON parse, 4 ms on GH200.
Mark Marosi (Mapika) released decider-2b, a 1.9B Qwen3.5-2B-Base fine-tune that does not generate text.
You pass a context plus typed questions with explicit option lists. One forward pass returns a probability for each option. No decoding, no JSON parse, no schema violations. The wire format matches TypeSafe Jev (POST /v1/systemone); their SDK works if TYPESAFE_BASE_URL points at decider.serve.
On a GH200 with CUDA graphs: 4.0 ms per request, about 1670 decisions/s in FP8. Choice (up to 255 options), boolean, and score fields, each with a calibrated confidence. English only, no reasoning — a fast pattern-matching classifier for routing, intent, and moderation.
Weights: Mapika/decider-2b. Code and training scripts: the Destination repo.
Run it
pip install git+https://github.com/Mapika/decider
from decider.infer import Decider
d = Decider("Mapika/decider-2b") # one CUDA GPU, ~4 GB
d.decide("My card was charged twice.",
[{"question": "Which team?", "options": ["billing", "technical", "sales"]}])
Needs torch, transformers>=5, and a CUDA GPU. Python 3.11+ recommended.