Agr
Command Code's open decision model, a Gemma-4-31B-IT fine-tune. It answers Choice, Score and Noul questions about a state in one forward pass, reading a probability for every option from the model's own output layer, and serves the System One request format.
Each question's block reads the shared state but never another question, so an answer does not depend on which other questions come with it. Option probabilities come from the backbone's own output layer over single-token option codes, with a temperature fitted on Command Code's development data; no calibration figure is published. Changing the option order can change the answer, and inputs are English text only. Command Code reports 58.15 on the third-party Decision Index 0.2.1 from its own run of the public kit (all 150,759 requests); Agr was not on the public board as of 6 October. The bf16 weights are 61.43 GB, served on 96 GB GPUs. Sibling: Agr-flash, SmolLM2-360M with a small trained scoring head (0.73 GB, 6,144-token state limit), which the maker says is not meant for safety decisions and publishes no score for. The agr package serves /v1/systemone; Command Code says it is not affiliated with TypeSafe, and credits Mark Marosi's Decider for the answer-slot prompt layout.
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
At a glance
| Parameters | 31B |
| Base model | google/gemma-4-31B-it |
| Maker | Command Code |
| Released | 2026-10-03 |
| License | apache-2.0 |
Get the weights
pip install systemonemodels
systemone pull command-code/agr
The files are served from the maker's Hugging Face repository, CommandCode/agr, and verified against the checksums recorded here.