Decision model
KnowLine 4B Gen5.5
1 question per request
- Accuracy, all 2,198 questions
- 48.5%46.4% to 50.6%, 95% interval; 1,066 right
- Answers kept at 99% confidence
- 39 of 2,1980 wrong; share right 91.0% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- No list priceBasis and price under Cost
- Latency
- 0.099 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- KnowLine server (knowline_server.py over SGLang 0.5.21, PelaAI's release code), self-hosted, FP8 at load
- Paid as
- Our own hardware
- Model ID
PelaAI/KnowLine-4B-Gen5.5- 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, PelaAI/KnowLine-4B-Gen5.5@779944d, huggingface.co/PelaAI/KnowLine-4B-Gen5.5; read 10 November 2026
- Hardware
- 1 × DGX Spark, GB10 (128 GB unified memory); FP8 at load
- Note
- PelaAI's released KnowLine-4B-Gen5.5 (PelaAI/KnowLine-4B-Gen5.5 at 779944d), a merged LoRA fine-tune of Qwen3.5-4B (the average of two) whose weights are those of Gen5. It ran on our hardware (one DGX Spark), one request at a time, through the card's serve_knowline.sh (SGLang 0.5.21 with the weights quantised to FP8 at load, and the repository's knowline_server.py, which reads a request of up to 3,000 tokens twice), with vision on. Its card lists English and Chinese.
- Confidence from
- Option score
- Run
- 965, 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
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 48.5% | 46.4% to 50.6% | 1,066 of 2,198 |
| Reading | 74.8% | 71.4% to 78.0% | 491 of 656 |
| Grammar | 37.3% | 34.9% to 39.7% | 575 of 1,542 |
Comparison group: δζηθ. Compare with another setup.
Accuracy by topic
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.
| Model | A · text onlyaccuracy, higher is better | B · image onlyaccuracy, higher is better | C · text plus imageaccuracy, higher is better | Best OCR routeaccuracy, higher is better |
|---|---|---|---|---|
| KnowLine 4B Gen5.5 Open · 4.0B | 45.7% | 30.1% | 45.3% | 39.1%via Surya 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
- 442 of 547, first English run (961)
- Accuracy
- 80.8%
- Non-hallucination with “I don't know”
- 0.9%: 1 “I don't know” of 106 not right (run 963)
- Answered with “I don't know”
- 99.8%: 546 of 547
- Error rate with “I don't know”
- 19.2%: 105 of 546 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- No list price
- Latency, median request, first English run
- 0.099 s; 1 request at once
- English run
- 961
- 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
- 39, of which 0 wrong (share right: 91.0% to 100.0%, 95% interval); 2,159 skipped, the model less than 99% sure
- Calibration gap
- 13.5 points
- AUROC, all questions
- 0.75 0.73 to 0.77
- AUROC, reading
- 0.82 0.79 to 0.85
- AUROC, grammar
- 0.64 0.62 to 0.67
- Brier score
- 0.220
- Accuracy of its surest answers
- most confident fifth: 87.7% right; most confident half: 66.4% right; most confident four fifths: 54.2% right
- Typical score of a right answer
- 80.1%
- Typical score of a wrong answer
- 47.9%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.099 s (this run)
- Latency, 90th percentile
- 0.145 s
- Requests at once
- 1
- Throughput
- 571.4 questions per minute of request time
- Request time
- 3.8 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
- 768
- Output tokens per question
- 1