OpenDecider-nano
Open-weights calibrated decision models by Manjunath Janardhan. The ~400M nano, an Ettin encoder with a decision head, answers typed choice, score and yes/no questions with a probability per option in one forward pass. 4B to 80B LoRA siblings. Apache-2.0.
A personal project. Each option gets its own marker, the logits are softmaxed per question, and all questions run in one batched pass; the answer space is set at request time. Siblings small, small-td (Qwen3-4B-Instruct-2507), medium-td (Qwen3-30B-A3B) and large-td (Qwen3-Next-80B-A3B) are LoRA adapters that read lettered next-token probabilities (Jev-inspired, like Tev1) and need the Qwen base weights. Nano and the -td models were fine-tuned on the typed-decisions train split, so that test score is in-distribution. "Calibrated" is the maker's claim; ECE was measured on 200 general decisions only (nano 0.092), with single training seeds. English only. The opendecider pip package includes a server compatible with Jev's /v1/systemone. The NOTICE lists permissively licensed training data and teacher models.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 400M |
| Base model | jhu-clsp/ettin-encoder-400m |
| Maker | Manjunath Janardhan |
| Released | 2026-09-27 |
| License | apache-2.0 |
| Reported accuracy | 79.6% |
| Reported latency | 17 ms per question on an NVIDIA L40S, 18 ms on an Apple M4 Max |
Get the weights
pip install systemonemodels
systemone pull manjunath-janardhan/opendecider
The files are served from the maker's Hugging Face repository, manjunathshiva/opendecider-nano, and verified against the checksums recorded here.