We build Decisio, an open-source server for the small decisions software asks a language model to make.
You send a piece of text and a typed question: yes or no, pick one option, or rate on a scale. Decisio answers with a probability for every option, read from the model in one forward pass. It generates no text, so there is nothing to parse, and "how sure is it" is a number you can act on.
It runs on your own hardware: a GPU server or a Mac. The base models are official open checkpoints, frozen and pinned. The tuned Gemma 4 12B adds an adapter trained for decisions, merged into its weights.
| Repository | Base | Runs on |
|---|---|---|
| decisio-gemma-4-31b | Gemma 4 31B, stored as FP8 (the default) | one 96 GB NVIDIA card |
| decisio-qwen3.6-35b-a3b | Qwen3.6-35B-A3B, FP8 | one 96 GB NVIDIA card |
| decisio-gemma-4-12b | Gemma 4 12B, bf16 | machines without a 96 GB card, including a Mac, and a laptop through Ollama 0.35.1 to 0.40.3 |
Right more often on hard, rule-heavy decisions. Against the untrained 12B in the same layout, it gains 6.7 points on 525 held-out hard decisions and 2.9 points on 939 decisions in everyday wording, and is about the same on a mixed set. All repositories below share the same tuned model; they differ in format and settings.
| Repository | Format | Runs on | Settings |
|---|---|---|---|
| Decisio-12b-tuned-calibrated | bf16 | one GPU (24 GB of weights) | honest probabilities |
| Decisio-12b-tuned-commit | bf16 | one GPU | commits yes/no answers |
| Decisio-12b-tuned-commit-think | bf16 | one GPU | commit, plus a short reasoning pass on near-tie choice questions; over-long inputs are cut instead of refused (decisio 0.35.0 or later) |
| Decisio-12b-tuned-fp8-calibrated / -fp8-commit | FP8 | a GPU with FP8 support | level with bf16 on accuracy |
| Decisio-12b-tuned-mlx8-calibrated / -mlx8-commit | MLX 8-bit | a Mac with 32 GB | the closest to bf16 |
| Decisio-12b-tuned-mlx-calibrated / -mlx-commit | MLX 6-bit | a Mac with 32 GB | the smaller download |
Each repository is pinned by a Decisio release, and the server refuses to start if a model's settings do not match the installed version.
On a GPU:
pip install "decisio[serve]"
python -m decisio.serve.vllm_engine
python -m decisio.serve.vllm_engine --base tachara-ai/Decisio-12b-tuned-calibrated
On a Mac:
pip install "decisio[mlx]"
python -m decisio.serve.vllm_engine --backend mlx --base tachara-ai/Decisio-12b-tuned-mlx8-calibrated
Code, documentation and the evaluation card: github.com/aminry/decisio. Website and the tuned 12B announcement: opendecisio.com.
Apache-2.0.