Decision model

KnowLine-0.8B-Gen1

1 question per request

Accuracy, all 2,198 questions
27.1%25.3% to 29.0%, 95% interval; 596 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
0.033 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-0.8B-Gen1
Parameters
800M dense, as the maker states it (counted from the released files: 0.87B).language model (tied embeddings) + multi-token prediction layer + 0.10B vision encoder; source: counted from the checkpoint's safetensors, PelaAI/KnowLine-0.8B-Gen1@6b27831, huggingface.co/PelaAI/KnowLine-0.8B-Gen1; read 11 November 2026
Hardware
1 × DGX Spark, GB10 (128 GB unified memory); BF16
Note
PelaAI's released KnowLine-0.8B-Gen1 (PelaAI/KnowLine-0.8B-Gen1 at 6b27831), a merged LoRA fine-tune of Qwen3.5-0.8B with its readout temperature (T = 1.54) folded into the weights. 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 vision tower is the base model's, not trained on images. Its card lists English and Chinese.
Confidence from
Option score
Run
1005, 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
e8e496050fe6
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions27.1%25.3% to 29.0%596 of 2,198
Reading38.3%34.6% to 42.0%251 of 656
Grammar22.4%20.4% to 24.5%345 of 1,542

Comparison group: ψΓΔ. Compare with another setup.

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading31.3%41 of 131
Sentence meaningsReading41.6%64 of 154
How a writer explains and develops ideasReading35.7%25 of 70
Topic and main point of a paragraphReading54.1%53 of 98
How a paragraph fits togetherReading30.2%32 of 106
Details that support a paragraph’s main pointReading37.1%36 of 97
Word types, such as nouns and verbsGrammar22.9%32 of 140
Groups of words built around a nounGrammar13.8%9 of 65
Verbs and their formsGrammar24.2%22 of 91
Verb forms used like nouns or modifiersGrammar11.1%8 of 72
Word endings and what they doGrammar21.5%17 of 79
How words are builtGrammar27.5%19 of 69
The roles of words and phrases in a sentenceGrammar22.8%33 of 145
How verbs show who acts and who receives the actionGrammar23.0%17 of 74
Types of sentencesGrammar22.4%19 of 85
Questions combining several grammar topicsGrammar15.9%35 of 220
Sound patterns and changesGrammar20.8%16 of 77
SpellingGrammar21.8%29 of 133
PunctuationGrammar33.3%39 of 117
Problems in sentence wordingGrammar28.6%50 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-0.8B-Gen1 Open · 0.8B24.5%23.9%24.1%22.8%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
322 of 547, first English run (1001)
Accuracy
58.9%
Non-hallucination with “I don't know”
10.7%: 26 “I don't know” of 243 not right (run 1003)
Answered with “I don't know”
95.2%: 521 of 547
Error rate with “I don't know”
41.7%: 217 of 521 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
No list price
Latency, median request, first English run
0.033 s; 1 request at once
English run
1001
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
8.3 points
AUROC, all questions
0.63 0.60 to 0.66
AUROC, reading
0.70 0.65 to 0.74
AUROC, grammar
0.57 0.53 to 0.60
Brier score
0.191
Accuracy of its surest answers
most confident fifth: 47.3% right; most confident half: 34.5% right; most confident four fifths: 28.8% right
Typical score of a right answer
35.1%
Typical score of a wrong answer
30.7%

Repeats

We did not repeat this setup.

Latency

Latency, median request
0.033 s (this run)
Latency, 90th percentile
0.047 s
Requests at once
1
Throughput
1702.2 questions per minute of request time
Request time
1.3 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