# StartLux releases Decision v1.1, from 0.8B to 27B, with non-commercial weights

> StartLux-Decision-v1.1 ships Instant and AutoThink variants. Weights are CC BY-NC 4.0; commercial use needs a separate licence.

Published 2026-10-11

Canonical: https://shipped.anth.us/decisions/startlux-decision-v1-1/

StartLux published StartLux-Decision v1.1 on Hugging Face in October 2026. The 27B Instant repository was created on 11 October, which is the date used here. The card describes a decision model that takes a state of text or JSON, with up to 8 images, plus typed questions: choice, yes/no and score. It returns a probability for every option in one forward pass. Its request format is compatible with the System One /v1/systemone API.

The family has 0.8B, 2B, 4B and 9B Instant models, 4B, 9B and 27B AutoThink models, and this 27B Instant model. Formats include BF16, GGUF and MLX.

The weights are CC BY-NC 4.0. Commercial use needs a separate licence from StartLux. The inference package is Apache-2.0.

The card's own runs give Decision Index 0.3 public subset 72.83 for 27B Instant against 73.37 for 27B AutoThink. It says these scores were used for model selection, so they are not independent test results. Median read-out latency on one H200 is about 52 ms at 256 tokens.

Engines that read the prompt causally serve a different model, the card warns. Another sign that the format is easy to copy and easy to get wrong.

## Sources
- https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-Instant

## Editor's first pass (AI-assisted)
- Real 3.5: Weights, GGUF and MLX builds are public, but non-commercial only.
- Useful 3.5: Same typed-decision API as Jev, runnable locally.
- Notable 3: The base for at least one other family on the Hub.
- Novel 3: Format follows Jev; Instant versus AutoThink is the twist.