Decision model
StartLux-Decision-v1.1-4B-Instant
1 question per request
- Accuracy, all 2,198 questions
- 47.9%45.8% to 50.0%, 95% interval; 1,052 right
- Answers kept at 99% confidence
- 9 of 2,1980 wrong; share right 70.1% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- No list priceBasis and price under Cost
- Latency
- 0.074 smedian request, 1 question per request
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-v1.1-4B-Instant- Parameters
- 4B dense, as the maker states it (counted from the released files: 4.66B).language model (tied embeddings) + multi-token prediction layer + 0.33B vision encoder; source: counted from the checkpoint's safetensors, startlux-models/StartLux-Decision-v1.1-4B-Instant@416088f, huggingface.co/startlux-models/StartLux-Decision-v1.1-4B-Instant; read 10 November 2026
- Hardware
- 1 × DGX Spark, GB10 (128 GB unified memory); BF16
- Note
- StartLux Labs' released StartLux-Decision-v1.1-4B-Instant (at 416088f), a fine-tune of Qwen3.5-4B 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
- 1236, 15 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
c2d81da72334- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 47.9% | 45.8% to 50.0% | 1,052 of 2,198 |
| Reading | 74.1% | 70.6% to 77.3% | 486 of 656 |
| Grammar | 36.7% | 34.3% to 39.1% | 566 of 1,542 |
Comparison group: ζηθ. Compare with another setup.
Accuracy by topic
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
- 453 of 547, first English run (1081)
- Accuracy
- 82.8%
- Non-hallucination with “I don't know”
- 0.0%: 0 “I don't know” of 94 not right (run 1083)
- Answered with “I don't know”
- 100.0%: 547 of 547
- Error rate with “I don't know”
- 17.2%: 94 of 547 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- No list price
- Latency, median request, first English run
- 0.075 s; 1 request at once
- English run
- 1081
- Machine, English runs
- DGX Spark
- Date
- 13 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 9, of which 0 wrong (share right: 70.1% to 100.0%, 95% interval); 2,189 skipped, the model less than 99% sure
- Calibration gap
- 7.1 points
- AUROC, all questions
- 0.76 0.74 to 0.78
- AUROC, reading
- 0.83 0.80 to 0.86
- AUROC, grammar
- 0.67 0.64 to 0.70
- Brier score
- 0.199
- Accuracy of its surest answers
- most confident fifth: 88.6% right; most confident half: 67.2% right; most confident four fifths: 54.3% right
- Typical score of a right answer
- 65.7%
- Typical score of a wrong answer
- 40.3%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.074 s (this run)
- Latency, 90th percentile
- 0.089 s
- Requests at once
- 1
- Throughput
- 773.5 questions per minute of request time
- Request time
- 2.8 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