Open weights · v1.1 · October 2026
Decides in one pass.
Thinks only when unsure.
StartLux-Decision models put a calibrated probability on every option of a typed question. Instant answers in a single forward pass. AutoThink answers the same way when it is sure, and writes a short thought first when it is not.
Decision Index 0.3, 27B AutoThink
DecisionPivot pair accuracy
median request, 27B Instant on one H200
open models, 0.8B to 27B
01Read-or-Think
Spend thought only where it is needed.
Every decision starts with one read-out: a probability for each option. When the top probability falls below the gate, AutoThink writes a short thought and reads again, and the answer blends both read-outs. Drag the gate.
“I was charged twice for order #4411 and the app still shows it as unpaid.”
Which team should handle this?
First read-out
Should this be answered today?
0.25 × first + 0.75 × second read-out
How severe is the problem?
0.25 × first + 0.75 × second read-out
Illustrative distribution and probabilities.
02Open weights
Every size, every build, one interface.
Each star is a repository on Hugging Face, sized by its all-time downloads and placed by parameter count. Instant glows amber, AutoThink blue, v1 ivory. Rings are MLX builds.
03Decision Index 0.3
Fits on one GPU. Ranked first of 174.
On 16 October 2026, 27B Instant ranked first of 174 entries on the Decision Index 0.3.2 leaderboard, and first on each of its three parts. The dashed filaments show what a thought adds at each size; it matters most where one detail decides the answer.
Pair accuracy on 7,800 contrast pairs that differ in one decisive detail.
AutoThink at gate 0.9. Source: technical report, arXiv 2610.17309, and the StartLux-Decision repository.
04Quick start
Serve a model in three commands.
Every model fits on one GPU. The 27B models run on one H200 in BF16 and on a 64 GB Apple laptop in MLX builds; the 0.8B and 2B models run on a CPU with llama.cpp.
$ git clone https://github.com/StartLuxLabs/StartLux-Decision.git && cd StartLux-Decision$ pip install -e '.[vllm]'$ hf download startlux-models/StartLux-Decision-v1.1-4B-AutoThink --local-dir models/4B-AutoThink$ startlux-decide serve --model models/4B-AutoThink --launch-vllm --gpus 0,1,2,3 --port 8090