Decider family
A family of open System One models by Mark Marosi that softmax option-letter logits at an answer slot: decider-0.8b, 2b, 4b and 35b-a3b on Qwen3.5 bases, decider-12b on Gemma-4-12B-it, and training-free readouts of larger chat models. This page carries decider-2b v11.
Unofficial listing, compiled from Mark Marosi's public releases; the source is the GitHub repository and the Hugging Face repositories linked below. The files here are decider-2b v11 (Qwen3.5-2B-Base, 1.9B parameters, 32k context), released on 24 September 2026 with decider-4b v2.1; the family's first release was on 16 September. The family: decider-0.8b (Qwen3.5-0.8B-Base), decider-2b v11, decider-4b v2.1 (Qwen3.5-4B-Base, 4.2B), decider-35b-a3b v1 (Qwen3.5-35B-A3B-Base, 34.7B total and 3B active, also in NVFP4), decider-2b-vision, and GGUF builds of the 2B and 4B, all Apache-2.0 and all serving POST /v1/systemone. On the author's regression set of 28 held-out tasks: decider-2b v11 75.2%, decider-4b v2.1 78.4%, decider-35b-a3b 81.0%. On the JevBench public hard tier: 57.7%, 64.9% and 67.6%. The author's Q8_0 GGUF build gave the same answer as the bf16 weights on 99.16% of 144,000 questions in the author's tests; the playground runs it on a CPU. On 29 September the author added decider-12b (google/gemma-4-12B-it with a merged rank-32 LoRA trained on generated state-tracking decisions; the untrained v1 stays under the tag v1). The author reports 0.856 argmax accuracy (ECE 0.144) on the test half of a fresh 399-item yes/no set and 0.847 (ECE 0.015) on 600 human- labelled Choice rows; on the JevBench public hard tier, in the author's own runs, v2 scored 0.712 against v1's 0.730. decider-chat-gemma4-31b and decider-chat-qwen3.6-27b involve no training: they are the stock Google and Qwen weights read through the decider readout with a fitted temperature, so this page carries no files for them. On the third-party Decision Index (v0.2.1, 28 September, measured by the index) they scored 57.33 and 51.35. decider-2b-coherent v1.1 (27 September) adds 9.5M parameters of add-on weights to a frozen decider-2b so that questions about one state get a single joint distribution; it runs on its own code rather than POST /v1/systemone. All are Apache-2.0.
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 | 1.9B |