ENPT

OpenDecisions

Last updated:

RuntimeApache-2.00

⚠️ Why not a model

Server and router. Serves the OneJev weights (catalogued separately) and can fall back to hosted Jev; it has no decision weights of its own.

Technical specs

Base LLMs
OneJev-4B (default), OneJev 0.8B / 9B / 27B, Jev (hosted, via TypeSafe or OpenRouter)
Decision types
choice, noul, score
Features
local, decisions-endpoint, systemone-endpoint, typesafe-sdk-compatible, gguf, multimodal, hosted-fallback, cli
License
Apache-2.0

Links

Not to be confused with OpenDecision (deepanwadhwa’s zero-shot engine) or the OpenDecider model.

OpenDecisions is an Apache-2.0 server published on 29 September 2026 (16:32 BRT) by the same maintainer as OneJev. It wraps the OneJev weights in one HTTP server and CLI: send text, images or video with a few fixed-option questions, and get back the chosen option and a probability for every option.

opendecisions serve            # OneJev-4B on PyTorch (CUDA) or llama.cpp
opendecisions ask "I was charged twice..." -q "Which team?" -o billing,tech,sales

What it does

  • Two endpoints. POST /v1/decisions uses its own shape (context, a list of questions with id, question, options, optional ordered: true for scales). POST /v1/systemone speaks the System One format, so the TypeSafe SDK works with TYPESAFE_BASE_URL pointed at the server.
  • Local model. OneJev-4B by default; other OneJev sizes can be selected. PyTorch when a CUDA GPU is present, llama.cpp otherwise.
  • Hosted fallback. With model: auto and a confidence threshold, questions the local model is unsure about can go to Jev through TypeSafe or OpenRouter (--remote, or --no-local to skip the local model). Each result says which provider answered.
  • Python client with no dependencies beyond the standard library, plus image() and video(frames, fps) helpers.

Caveats

  • pip install opendecisions does not work yet. The README uses it, but the package is not on PyPI (404 on 01/10); install from the repo. The name is unclaimed.
  • One “init” commit and no stars as of 01/10.
  • The probabilities are the OneJev model’s; no extra calibration is added by the server.