Decision model

GLiClass v4 Edge

1 question per request

Accuracy, all 2,198 questions
23.3%21.6% to 25.2%, 95% interval; 513 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
0.025 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Decision model
Access
GLiClass server (our servers/gliclass_server.py: the card's gliclass pipeline and its System One mapping), self-hosted, BF16
Paid as
Our own hardware
Model ID
knowledgator/gliclass-v4-edge
Parameters
400M dense, as the maker states it (counted from the released files: 0.44B).DeBERTa-v3-large encoder + label projectors and scorer + calibrator (text only); source: counted from the checkpoint's safetensors, knowledgator/gliclass-v4-edge@885af1a, huggingface.co/knowledgator/gliclass-v4-edge; read 11 November 2026
Hardware
1 × DGX Spark, GB10 (128 GB unified memory); BF16
Note
Knowledgator's released GLiClass v4 Edge (knowledgator/gliclass-v4-edge at 885af1a), a zero-shot classifier fine-tuned from DeBERTa-v3-large (supervised training, reinforcement learning, then a calibrator). Its card serves it as a library pipeline (gliclass) and maps a decision request onto it: the passage is the text, the instructions the prompt, each option a label, and a softmax over the calibrated label scores. It ran on our hardware (one DGX Spark), one request at a time, through our HTTP server (servers/gliclass_server.py) with that mapping and gliclass 0.1.22, in BF16, text only; a request over 16,384 tokens (the card's limit) is refused, never cut. Its card lists English only.
Confidence from
Option score
Run
1046, 12 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
6ad0eac51a4a
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions23.3%21.6% to 25.2%513 of 2,198
Reading25.9%22.7% to 29.4%170 of 656
Grammar22.2%20.2% to 24.4%343 of 1,542

Comparison group: ΓΔΘΛ. Compare with another setup.

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading24.4%32 of 131
Sentence meaningsReading28.6%44 of 154
How a writer explains and develops ideasReading21.4%15 of 70
Topic and main point of a paragraphReading42.9%42 of 98
How a paragraph fits togetherReading20.8%22 of 106
Details that support a paragraph’s main pointReading15.5%15 of 97
Word types, such as nouns and verbsGrammar19.3%27 of 140
Groups of words built around a nounGrammar20.0%13 of 65
Verbs and their formsGrammar17.6%16 of 91
Verb forms used like nouns or modifiersGrammar22.2%16 of 72
Word endings and what they doGrammar25.3%20 of 79
How words are builtGrammar26.1%18 of 69
The roles of words and phrases in a sentenceGrammar27.6%40 of 145
How verbs show who acts and who receives the actionGrammar24.3%18 of 74
Types of sentencesGrammar24.7%21 of 85
Questions combining several grammar topicsGrammar23.2%51 of 220
Sound patterns and changesGrammar20.8%16 of 77
SpellingGrammar20.3%27 of 133
PunctuationGrammar20.5%24 of 117
Problems in sentence wordingGrammar20.6%36 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
GLiClass v4 Edge Open · 0.4B22.0%No image inputNo image input27.1%via Surya OCR 2, best of the 3 of 4 OCR routes measured (Baidu Unlimited-OCR-2: 762 of 763 questions answered)

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
319 of 547, first English run (1042)
Accuracy
58.3%
Non-hallucination with “I don't know”
21.0%: 52 “I don't know” of 248 not right (run 1044)
Answered with “I don't know”
90.5%: 495 of 547
Error rate with “I don't know”
39.6%: 196 of 495 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
No list price
Latency, median request, first English run
0.017 s; 1 request at once
English run
1042
Machine, English runs
DGX Spark
Date
12 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
10.5 points
AUROC, all questions
0.56 0.53 to 0.59
AUROC, reading
0.62 0.57 to 0.67
AUROC, grammar
0.52 0.49 to 0.56
Brier score
0.192
Accuracy of its surest answers
most confident fifth: 30.7% right; most confident half: 25.7% right; most confident four fifths: 24.4% right
Typical score of a right answer
32.3%
Typical score of a wrong answer
30.1%

Repeats

We did not repeat this setup.

Latency

Latency, median request
0.025 s (this run)
Latency, 90th percentile
0.042 s
Requests at once
1
Throughput
2104.6 questions per minute of request time
Request time
1.0 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
396
Output tokens per question
0