Decision model

KnowLine-31B-Gen1.5

1 question per request

Accuracy, all 2,198 questions
81.8%80.1% to 83.4%, 95% interval; 1,798 right
Answers kept at 99% confidence
143 of 2,1980 wrong; share right 97.4% to 100.0%, 95% interval
Cost, $ per 1,000 questionsper 1M questions
No list priceBasis and price under Cost
Latency
1.05 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Decision model
Access
KnowLine server (knowline_server.py over SGLang 0.5.21, PelaAI's release code), self-hosted, BF16
Paid as
Our own hardware
Model ID
PelaAI/KnowLine-31B-Gen1.5
Parameters
31B dense, as the maker states it (counted from the released files: 31.27B).language model (tied embeddings) + 0.58B vision encoder; source: counted from the checkpoint's safetensors, PelaAI/KnowLine-31B-Gen1.5@8e56aa1, huggingface.co/PelaAI/KnowLine-31B-Gen1.5; read 10 November 2026
Hardware
1 × DGX Spark, GB10 (128 GB unified memory); BF16
Note
PelaAI's released KnowLine-31B-Gen1.5 (PelaAI/KnowLine-31B-Gen1.5 at 8e56aa1), a merged LoRA fine-tune of Gemma-4-31B-it whose weights are those of Gen1; Gen1.5 changes only how long requests are served. It ran on our hardware (one DGX Spark), one request at a time, through the card's serve_knowline.sh (SGLang 0.5.21 in BF16 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
957, 10 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
57f3caac4692
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions81.8%80.1% to 83.4%1,798 of 2,198
Reading92.8%90.6% to 94.6%609 of 656
Grammar77.1%74.9% to 79.1%1,189 of 1,542

Comparison group: nop. Compare with another setup.

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading88.5%116 of 131
Sentence meaningsReading96.1%148 of 154
How a writer explains and develops ideasReading88.6%62 of 70
Topic and main point of a paragraphReading94.9%93 of 98
How a paragraph fits togetherReading88.7%94 of 106
Details that support a paragraph’s main pointReading99.0%96 of 97
Word types, such as nouns and verbsGrammar85.0%119 of 140
Groups of words built around a nounGrammar89.2%58 of 65
Verbs and their formsGrammar86.8%79 of 91
Verb forms used like nouns or modifiersGrammar93.1%67 of 72
Word endings and what they doGrammar82.3%65 of 79
How words are builtGrammar63.8%44 of 69
The roles of words and phrases in a sentenceGrammar75.9%110 of 145
How verbs show who acts and who receives the actionGrammar82.4%61 of 74
Types of sentencesGrammar88.2%75 of 85
Questions combining several grammar topicsGrammar66.4%146 of 220
Sound patterns and changesGrammar44.2%34 of 77
SpellingGrammar73.7%98 of 133
PunctuationGrammar74.4%87 of 117
Problems in sentence wordingGrammar83.4%146 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
KnowLine-31B-Gen1.5 Open · 31B79.0%49.8%76.0%75.5%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
512 of 547, first English run (953)
Accuracy
93.6%
Non-hallucination with “I don't know”
0.0%: 0 “I don't know” of 37 not right (run 955)
Answered with “I don't know”
100.0%: 547 of 547
Error rate with “I don't know”
6.8%: 37 of 547 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
No list price
Latency, median request, first English run
1.12 s; 1 request at once
English run
953
Machine, English runs
DGX Spark
Date
10 November 2026
Request format
1 question per request
Reasoning
off
Prompt version
47f5776e71db
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

Confidence

Answers kept at 99% or more
143, of which 0 wrong (share right: 97.4% to 100.0%, 95% interval); 2,055 skipped, the model less than 99% sure
Calibration gap
3.4 points
AUROC, all questions
0.87 0.85 to 0.88
AUROC, reading
0.93 0.91 to 0.96
AUROC, grammar
0.83 0.81 to 0.85
Brier score
0.109
Accuracy of its surest answers
most confident fifth: 100.0% right; most confident half: 98.3% right; most confident four fifths: 90.4% right
Typical score of a right answer
90.4%
Typical score of a wrong answer
54.8%

Repeats

We did not repeat this setup.

Latency

Latency, median request
1.05 s (this run)
Latency, 90th percentile
1.48 s
Requests at once
1
Throughput
53.7 questions per minute of request time
Request time
40.9 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
737
Output tokens per question
1