Decima-small 1.1
A 122M-parameter late-interaction decision model on multilingual-e5-small: choose, verify, ordered score and rank with a probability per option, int8 ONNX at about 20 ms on one CPU core, the same answer in any option order, and a local server in the System One wire format.
A personal release by Amir Mahdi Madani. The state and question are encoded once, each free-text option is encoded separately and cross-attends to the state, and the model gives one logit per option, so option order cannot change the answer. One shipped temperature, fitted on a hold-out of the training mix. On the community jabr/classifier-benchmark v2 (a third-party suite) the maker's own int8 run gives 0.616 macro accuracy; score is the weakest type. Ships int8 and fp32 ONNX exports and PyTorch weights; runs in the browser. Siblings decima-base and decima-agent.
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
- rank — orders a set of items
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 122M |
| Base model | intfloat/multilingual-e5-small |
| Maker | Amir Mahdi Madani |
| Released | 2026-09-27 |
| License | apache-2.0 |
| Reported accuracy | 61.6% |
| Reported latency | 20 ms p50 / 23 ms p95 on one Intel Core Ultra 7 P-core, ONNX Runtime int8, 4 options |
Get the weights
pip install systemonemodels
systemone pull amyrmahdy/decima
The files are served from the maker's Hugging Face repository, amyrmahdy/decima-small, and verified against the checksums recorded here.