StartLux has released StartLux-Decision v1.1, a family of open decision models, and its 27B Instant model now sits at the top of the Decision Index 0.3.2 leaderboard on Hugging Face. StartLux is an incubated company of Genora.
Decision Index is a community-maintained board and is not affiliated with TypeSafe. As of October 16, the 27B Instant model ranked first with a Full score of 69.12. It also led each of the three parts: 72.88 on the public benchmarks, 64.82 on private tests of the same skills, and 61.41 on private tasks from new domains. The runner-up, vLLM-SR d3, scored 66.26, and Jev 1.14 scored 61.53.
The models are built for a narrow, practical job. Given a state, which can be text, JSON or images, and a set of typed questions, a model returns a probability for every option: a pick from a list, a yes-or-no judgment, or a level on a scale. Instant models answer in a single forward pass, using bidirectional attention over the prompt. AutoThink models answer the same way when confident, but when the top probability falls below a gate, 0.6 by default, they write a short reasoning trace and then read the answer again.
The lineup covers eight open models from 0.8B to 27B parameters in Instant and AutoThink variants, with GGUF and MLX quantized formats. They run on vLLM and MLX. Weights are released under CC BY-NC 4.0, which bars commercial use.
On the public subset of Decision Index 0.3, StartLux reports a self-run score of 73.37 for the 27B AutoThink model at a gate of 0.9. The company also introduced DecisionPivot, its own benchmark of verifiable decisions built from contrast pairs that differ in a single answer-changing detail, covering policies, ledgers, schedules, permissions and tool calls. StartLux reports 86.9% pair accuracy for the 27B AutoThink model there. Both figures are company-reported rather than independent results.
The technical report, posted on arXiv as 2610.17309, describes a three-stage training pipeline and a reinforcement learning method called RLCD-squared. It rewards better-calibrated decision probabilities using proper scoring rules and softly penalizes overlong reasoning. The authors say it cut average reasoning length by up to about 65% while largely keeping accuracy.
Demos on the project site show the models making step-by-step calls in chess, Super Mario Bros., StarCraft II and a computer-use shopping task, with one typed request per step.
SEE ALSO: StartLux's 27B Local Model Beats DeepSeek V4 Flash in China AI Benchmark · Ant Group Open-Sources Ling-3.1-flash Weights Under MIT