ENPT

OpenDecider

Last updated:

Open9–18ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity6.1/10

    Apache-2.0 weights, PyPI package, Colab notebook, HF Space demo and a detailed COMPARISON.md. Self-host only; no hosted SLA. First public release 27 Sep 2026.

  • Capability8/10

    First-class choice / score / noul with published calibration talk and ~17 ms L40S / ~18 ms M4 Max claims. Strong author head-to-head tables with CIs; nano typed-decisions score is in-domain on the train split.

  • Adoption17/100

    Reddit MachineLearningNews thread (~51 upvotes), PyPI 0.1.2, about 10 GitHub stars and a live Space. Early ecosystem traction, no production case studies yet.

Vendor claims

OpenDecider is an open family of calibrated System One decision models. You pass a state (text, ticket or JSON) and typed questions (choice, score, noul); it returns a probability for every option in one forward pass. No text generation.

Default entry point: manjunathshiva/opendecider-nano (~400M, Ettin encoder). Siblings: opendecider-small (4B LoRA on Qwen3-4B-Instruct-2507), workflow-tuned *-td variants, and opendecider-medium-td (Qwen3-30B-A3B + LoRA). Install: pip install opendecider (PyPI 0.1.2). Code and tables: github.com/manjunathshiva/opendecider (~10★). Live demo: HF Space.

Independent of TypeSafe AI. Distinct from the earlier watch item mvbalaji/od1* (“Open Decider”) — different author and stack.

Specs

Attribute Value
Author manjunathshiva (independent)
Flagship opendecider-nano (~400M)
Other public weights small, small-td, medium-td, MLX 4/8-bit small builds
Decision types choice, score, noul (catalog boolean for noul)
License Apache-2.0
Released
Runtime opendecider Python package — CPU, NVIDIA, Apple Silicon (MPS/MLX)

Author benchmarks

Everything here is author-reported → UnverifiedClaim, scored with the Antz AI / jev_and_laya_benchmarking harness and TypeSafe’s API for the Jev row. Full rebuild notes: COMPARISON.md.

Model typed-decisions Note
opendecider-nano 0.796 Fine-tuned on train split (like Laya’s TD checkpoint)
Laya TD checkpoint 0.766 Same harness
TypeSafe Jev 0.754 Via TypeSafe API
opendecider-small 0.672 Never saw typed-decisions
opendecider-small-td 0.792 Workflow fine-tune
opendecider-medium-td 0.788 Also 0.765 on 200 general decisions (author)

Read the in-domain column first for nano / *-td. Prefer small (no TD fine-tune) when you need a cleaner zero-shot comparison.

Limits

  • Nano’s headline typed-decisions number is in-domain. Same caveat as Laya’s TD checkpoint and RSI-Jev suite scores.
  • Self-run head-to-heads. Careful CIs, but not an independent leaderboard entry.
  • Medium-td is NVIDIA/Linux-oriented in the README.
  • Not TypeSafe. system_one() naming is an API convenience, not affiliation.

Fit / anti-fit

Fit when you need: a pip-installable open decision family with CPU/Mac/NVIDIA parity claims, Colab/Space to try first, or a small encoder peer next to Julia-1 / Laya.

Anti-fit when you need: multimodal vision (see Valen / Jev-Omni), a hosted SLA, or a number you can treat as zero-shot without reading which checkpoint was TD-tuned.

Why it is in the catalog

Trained decision weights with a public typed interface, Apache-2.0 distribution, and documented (if self-run) comparisons against Jev and Laya. Family card, nano as default — same pattern as JevEmbed / RSI-Jev.