StartLuxLabs/StartLux-Decision

What it claims

StartLux-Decision v1.1 27B Instant ranked first of 174 entries on the Decision Index 0.3.2 leaderboard on Hugging Face (October 16, 2026). It was also first on all three parts: public benchmarks (72.88), private tests of the same skills (64.82) and private tasks from new domains (61.41). The repository says the Instant models beat the strongest evaluated decision model of comparable size by 7.4 to 16.9 points on Decision Index 0.3. AutoThink adds 11.5 to 18.4 points of DecisionPivot pair accuracy over Instant. The technical report is arXiv 2610.17309 (October 2026).

How it works

Instant models answer in one forward pass with bidirectional attention over the prompt. AutoThink models also read out answers in one pass first. If the top probability is below a gate (0.6 by default), the model writes a short reasoning trace and then reads the answer again. AutoThink reasoning is trained with RLCD², described as reinforcement learning for calibrated deliberative decisions. The page does not give training data or compute.

flowchart LR
  A[State + typed questions] --> B[One-pass read-out]
  B --> C{Top prob >= gate?}
  C -- yes --> D[Return probabilities]
  C -- no --> E[Write reasoning trace]
  E --> F[Second read-out]
  F --> D

Evidence

The gate choice matters most on rule and arithmetic questions. For the 27B AutoThink model, moving the gate from 0.9 to 0.6 lowers Decision Index 0.3 only slightly, from 73.22 to 72.85. DecisionPivot pair accuracy falls sharply, from 88.1% to 68.4%. DecisionPivot has 7,800 generated examples in contrast pairs that differ in one decisive detail, and a pair counts only if both answers are correct.

Latency was measured on one H200 with vLLM and one request in flight. Instant models take 25 to 52 ms at the median, from 0.8B to 27B. A confident AutoThink read-out takes 38 to 59 ms, and an unsure one takes longer as its thought grows.

For quantization, FP8 and GGUF Q8_0 largely preserve BF16 performance. Q4_K_M can lose up to 4.3 points (27B AutoThink on DecisionPivot).

How to use it

Install into the same Python environment as vLLM. The package targets vLLM 0.31.0. Published numbers were measured on 0.19.1, which is also supported.

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/StartLux-Decision-v1.1-4B-AutoThink
startlux-decide serve --model models/StartLux-Decision-v1.1-4B-AutoThink --launch-vllm --gpus 0,1,2,3 --port 8090

A request to POST /v1/systemone has a state object and named questions. Each question has a type (choice, noul or score), instructions, and criteria: a map of option to description for choice, or an ordered list of levels for score. Choice answers return probabilities, confidence and none_probability. noul returns the probability of yes. score returns the expected level plus a probability per level. The response also reports token usage.

The README suggests treating none_probability of 0.2 or more as "no listed option fits". At that threshold it catches 66.3% of questions whose options are all wrong and flags 2.2% of questions that have a right answer.

startlux-decide also has bench (latency), eval (accuracy and calibration on labelled requests) and quantize (FP8, GGUF, MLX). The server supports API keys, HTTPS, early 429 responses and access logs for public serving. Tests run with python -m unittest -v using only the standard library.

Without a GPU, bamboonode.dev hosts the 27B AutoThink model free and without usage limits at POST /v1/decisions, with the same request format. It needs a Bamboo API key.

Limits

The weights are CC BY-NC 4.0, so commercial use requires a licence from StartLux. The code is Apache 2.0. Published AutoThink results use a gate of 0.9, while the server default is 0.6. At 0.6, DecisionPivot accuracy is about 20 points lower than reported. AutoThink latency is unbounded by the page's figures when the model reasons. Decision Index 0.3 numbers are on the public subset with the public kit. The v1 models answer only in one pass and have no AutoThink variant.