StartLux opens eight small decision models from 0.8B to 27B

A product only needs a yes or no from a model. It still pays for a full chat call to get the answer. StartLux released a family of eight open decision models today that answer a short question in one step. The repository is called StartLux-Decision. It has 1,473 stars on GitHub and was created on 29 September.

The models run from 0.8B to 27B parameters. Each one can answer three kinds of question. It picks one option from a list. It gives a yes or no. It rates something on a scale. The output is a probability for every option in one forward pass.

The eight models

Four model sizes come in two variants. Instant answers in one pass. AutoThink writes a short thought first when the model is unsure, then answers. The 0.8B and 2B sizes are Instant only. A think budget of 4,096 tokens caps the longest thought AutoThink will produce.

All eight weights are on Hugging Face under the startlux-models organisation.

StartLux's benchmark numbers

StartLux states its own benchmark numbers. The test setup is one H200 GPU running vLLM with one request in flight. On its Decision Index 0.3, the 27B AutoThink model scores 73.37 out of 100. The 27B Instant model scores 72.83. The 0.8B Instant model scores 53.75.

On a 7,800-pair test called DecisionPivot, the 27B AutoThink pair accuracy is 86.9 percent. The 0.8B drops to 8.1 percent on the same test. The gap between the smallest and the largest model is wide, so test the size that fits the task.

Instant models answer at 25 to 52 milliseconds at the median across the full size range. AutoThink answers in 38 to 59 milliseconds when it skips the reasoning step. These figures are StartLux's own and have not been reproduced on independent hardware.

Where it runs and the licence

The weights work with vLLM, llama.cpp, MLX for Apple Silicon, and the Hugging Face Transformers library. StartLux also provides its own engine called StartLux-Prism for NVIDIA GPUs. The code is Apache 2.0. The weights are CC BY-NC 4.0, which forbids commercial use without a paid licence from StartLux.

For a team routing short classifier calls through a chat model, this is a cheaper option worth a side-by-side test. Many classifier calls on a product are decision shapes. Is this message a refund request? Does this ticket need a human? Rate this risk from one to five. A 2B decision model runs one of those in 25 to 52 milliseconds. The money saved on chat calls can pay for harder work that needs reasoning. Read the licence first if you plan to release a paid product on these weights.

Source

StartLuxLabs/StartLux-Decision on GitHub.