// PROOF · DECIMA
A 122M-parameter decision model that never changes its answer when you reorder the options
Decima is a small, calibrated, multilingual decision model — situation, question and free-text options in; one probability per option out — that runs in about 20 ms on a single laptop CPU core and is released with open weights under Apache-2.0.
METHODEvery competitor was run from its own checkpoint on one harness, after reproducing its published numbers exactly with its own evaluation code. Latency is the median of 40 timed calls on an Intel Core Ultra 7 155H P-core, ONNX Runtime, option encodings cached. The shuffle test re-asks every multi-option question with the options permuted. Decima trained on the MASSIVE and XNLI train splits; Laya reports it did not — the 0.764 is in-distribution for Decima and zero-shot for Laya. Full protocol, overlap audit and limitations are in the evaluation document.
Most AI decisions inside a product are not questions. They are choices. Which team gets this ticket. Is this message about billing. How urgent is this incident. The options are known before the request arrives; only the judgement is missing.
The industry answers these with a large language model, a prompt, and a JSON parser — seconds of latency and a cloud bill for a decision a person makes in a glance. Jev showed that a model built to decide rather than to write can do better. Decima asks how small such a model can be.
The architectural decision came first. The situation is encoded once; every option is encoded on its own and reads the situation through cross-attention. No option ever sees another option or its own position. Two consequences follow, and neither is a tuning result. Reordering the options cannot change the answer — the open alternatives we measured change theirs 10 to 27 per cent of the time. And there is no prompt to overflow: an option set of four and one of a thousand run through the same path, with option encodings cached so a repeated question pays only for the new situation.
The rest is disciplined training: a 26-billion-parameter teacher labelling three hundred thousand decisions with full probability distributions across four languages, gold labels from public datasets, and a model-selection score that counted zero-shot ability, so gains on familiar benchmarks could not hide a loss in generalisation. When plain int8 quantisation cost too much — its top answer matched full precision on only 88 per cent of real test items — block-wise 8-bit brought agreement to 99 per cent at a quarter of the size.
It is not a general model. Long, multi-fact business decisions and questions that need world knowledge remain beyond it, and the evaluation says so on the first page.
A decision that software can act on needs two things: the choice, and an honest measure of how sure it is. Everything else is latency.
Model — huggingface.co/amyrmahdy/decima-small · Try it — Decima Playground · Code & technical report — github.com/amyrmahdy/decima
PyTorchONNX Runtimemultilingual-E5Gemma 4 (teacher)vLLMGradio