Decision model

StartLux-Decision-v1.1-2B-Instant

1 question per request

Accuracy, all 2,198 questions
34.3%32.3% to 36.3%, 95% interval; 753 right
Answers kept at 99% confidence
No answers at this thresholdnone of the 2,198 questions at 99% or more
Cost, $ per 1,000 questionsper 1M questions
No list priceBasis and price under Cost
Latency
0.040 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Decision model
Access
StartLux-Decision server (startlux-decide on vLLM 0.31.0, the maker's release code), self-hosted, BF16
Paid as
Our own hardware
Model ID
startlux-models/StartLux-Decision-v1.1-2B-Instant
Parameters
2B dense, as the maker states it (counted from the released files: 2.27B).language model (tied embeddings) + multi-token prediction layer + 0.33B vision encoder; source: counted from the checkpoint's safetensors, startlux-models/StartLux-Decision-v1.1-2B-Instant@f1e6cd1, huggingface.co/startlux-models/StartLux-Decision-v1.1-2B-Instant; read 11 November 2026
Hardware
1 × DGX Spark, GB10 (128 GB unified memory); BF16
Note
StartLux Labs' released StartLux-Decision-v1.1-2B-Instant (at f1e6cd1), a fine-tune of Qwen3.5-2B that reads the prompt bidirectionally and also reads images. It ran on our hardware (one DGX Spark), one request at a time, through the server of its package (startlux-decide serve --launch-vllm, StartLux-Decision at a99977d) on vLLM 0.31.0, with vision on: the card's scores were measured on vLLM 0.19.1, which has no build for the DGX Spark's processor (aarch64). Its weights are licensed for non-commercial use only (CC BY-NC 4.0); we use them non-commercially, for evaluation. Its card lists English only.
Confidence from
Option score
Run
977, 11 November 2026

Settings

Reasoning
off
Request format
1 question per request
Requests at once
1
Temperature
not recorded
Run record: harness seed, harness commit and prompt version
Harness seed
0
Harness commit
b55c22654bc7
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions34.3%32.3% to 36.3%753 of 2,198
Reading52.9%49.1% to 56.7%347 of 656
Grammar26.3%24.2% to 28.6%406 of 1,542

Comparison group: ξπ. Compare with another setup.

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading42.0%55 of 131
Sentence meaningsReading60.4%93 of 154
How a writer explains and develops ideasReading47.1%33 of 70
Topic and main point of a paragraphReading73.5%72 of 98
How a paragraph fits togetherReading41.5%44 of 106
Details that support a paragraph’s main pointReading51.5%50 of 97
Word types, such as nouns and verbsGrammar30.7%43 of 140
Groups of words built around a nounGrammar32.3%21 of 65
Verbs and their formsGrammar31.9%29 of 91
Verb forms used like nouns or modifiersGrammar30.6%22 of 72
Word endings and what they doGrammar31.6%25 of 79
How words are builtGrammar24.6%17 of 69
The roles of words and phrases in a sentenceGrammar24.1%35 of 145
How verbs show who acts and who receives the actionGrammar27.0%20 of 74
Types of sentencesGrammar25.9%22 of 85
Questions combining several grammar topicsGrammar24.5%54 of 220
Sound patterns and changesGrammar19.5%15 of 77
SpellingGrammar25.6%34 of 133
PunctuationGrammar17.9%21 of 117
Problems in sentence wordingGrammar27.4%48 of 175

Vision (preliminary)

Preliminary: these numbers are not final and may change. This setup is condition A of the Vision tab, which asks the 763 questions with underlined or numbered words as text (A), as a page image (B), as both (C) and through OCR models (D); the share answered right, its 95% interval in the number's tooltip.

ModelA · text onlyaccuracy, higher is betterB · image onlyaccuracy, higher is betterC · text plus imageaccuracy, higher is betterBest OCR routeaccuracy, higher is better
StartLux-Decision-v1.1-2B-Instant Open · 2.0B32.0%23.5%30.3%29.0%via Baidu Unlimited-OCR-2

English

The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.

Right answers
381 of 547, first English run (973)
Accuracy
69.7%
Non-hallucination with “I don't know”
16.9%: 32 “I don't know” of 189 not right (run 975)
Answered with “I don't know”
94.1%: 515 of 547
Error rate with “I don't know”
30.5%: 157 of 515 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
No list price
Latency, median request, first English run
0.040 s; 1 request at once
English run
973
Machine, English runs
DGX Spark
Date
11 November 2026
Request format
1 question per request
Reasoning
off
Prompt version
47f5776e71db
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

Confidence

Answers kept at 99% or more
No answers at this threshold; 2,198 skipped, the model less than 99% sure
Calibration gap
6.4 points
AUROC, all questions
0.66 0.64 to 0.69
AUROC, reading
0.72 0.68 to 0.76
AUROC, grammar
0.57 0.53 to 0.60
Brier score
0.204
Accuracy of its surest answers
most confident fifth: 60.1% right; most confident half: 43.8% right; most confident four fifths: 36.8% right
Typical score of a right answer
41.9%
Typical score of a wrong answer
32.9%

Repeats

We did not repeat this setup.

Latency

Latency, median request
0.040 s (this run)
Latency, 90th percentile
0.046 s
Requests at once
1
Throughput
1461.2 questions per minute of request time
Request time
1.5 minutes, over the requests whose answers were kept

Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.

Cost

Cost per 1,000 questionsper 1M questions
No list price
Basis
No dollar estimate; see hardware and time
Input tokens per question
363
Output tokens per question
0