Decision model

GLiDE 9B

1 question per request

Accuracy, all 2,198 questions
52.0%50.0% to 54.1%, 95% interval; 1,144 right
Answers kept at 99% confidence
27 of 2,1980 wrong; share right 87.5% to 100.0%, 95% interval
Cost, $ per 1,000 questionsper 1M questions
No list priceBasis and price under Cost
Latency
0.207 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Decision model
Access
GLiDE server (Fastino's release code), self-hosted
Paid as
Our own hardware
Model ID
glide-9b
Variant
local
Parameters
7.94B dense, as the maker states it (counted from the released files: 7.94B).language model (no LM head) + 255-way readout; source: counted from the checkpoint's safetensors, fastino/GLiDE-9B@ab54df5, huggingface.co/fastino/GLiDE-9B; read 10 November 2026
Hardware
1 × DGX Spark, GB10 (128 GB unified memory); BF16
Note
Fastino's released GLiDE 9B (fastino/GLiDE-9B at ab54df5), on Qwen3.5-9B, text only, whose layers 12 to 23 run twice (recurrent depth, its decision_config.json), run on our hardware through the server it ships.
Confidence from
Option score
Run
907, 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
3b14502cefed
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions52.0%50.0% to 54.1%1,144 of 2,198
Reading80.2%77.0% to 83.1%526 of 656
Grammar40.1%37.7% to 42.5%618 of 1,542

Comparison group: Zαβγδζ. Compare with another setup.

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading70.2%92 of 131
Sentence meaningsReading87.0%134 of 154
How a writer explains and develops ideasReading64.3%45 of 70
Topic and main point of a paragraphReading91.8%90 of 98
How a paragraph fits togetherReading73.6%78 of 106
Details that support a paragraph’s main pointReading89.7%87 of 97
Word types, such as nouns and verbsGrammar44.3%62 of 140
Groups of words built around a nounGrammar33.8%22 of 65
Verbs and their formsGrammar44.0%40 of 91
Verb forms used like nouns or modifiersGrammar44.4%32 of 72
Word endings and what they doGrammar50.6%40 of 79
How words are builtGrammar29.0%20 of 69
The roles of words and phrases in a sentenceGrammar37.9%55 of 145
How verbs show who acts and who receives the actionGrammar51.4%38 of 74
Types of sentencesGrammar30.6%26 of 85
Questions combining several grammar topicsGrammar30.0%66 of 220
Sound patterns and changesGrammar22.1%17 of 77
SpellingGrammar36.1%48 of 133
PunctuationGrammar44.4%52 of 117
Problems in sentence wordingGrammar57.1%100 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
GLiDE 9B Open · 7.9B48.2%No image inputNo image input44.2%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
472 of 547, first English run (909)
Accuracy
86.3%
Non-hallucination with “I don't know”
16.7%: 14 “I don't know” of 84 not right (run 911)
Answered with “I don't know”
97.4%: 533 of 547
Error rate with “I don't know”
13.1%: 70 of 533 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
No list price
Latency, median request, first English run
0.207 s; 1 request at once
English run
909
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
27, of which 0 wrong (share right: 87.5% to 100.0%, 95% interval); 2,171 skipped, the model less than 99% sure
Calibration gap
8.5 points
AUROC, all questions
0.79 0.77 to 0.81
AUROC, reading
0.88 0.85 to 0.91
AUROC, grammar
0.69 0.66 to 0.72
Brier score
0.191
Accuracy of its surest answers
most confident fifth: 96.1% right; most confident half: 73.2% right; most confident four fifths: 59.2% right
Typical score of a right answer
74.7%
Typical score of a wrong answer
44.9%

Repeats

We did not repeat this setup.

Latency

Latency, median request
0.207 s (this run)
Latency, 90th percentile
0.253 s
Requests at once
1
Throughput
281.0 questions per minute of request time
Request time
7.8 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
359
Output tokens per question
0

Going first in front of an LLM

Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.

LLM behind itShare takenAccuracy changeSaving
Gemini 3.8 Flash high · chat12.4%−0.14 points12%†
GPT-6 Astra high · chat12.0%−0.13 points12%†
GPT-6.1 Sol high · chat13.0%−0.16 points13%†
GPT-6 Astra low · chat12.0%−0.13 points12%†
GPT-6 Astra low · chat · 1 per request11.8%−0.11 points12%†
Gemini 3.8 Flash low · chat · 1 per request12.3%−0.15 points12%†
GPT-6.1 Sol low · chat · 1 per request12.8%−0.19 points13%†
GPT-6.1 Sol low · chat12.0%−0.13 points12%†
Claude Opus 5.5 high · chat13.3%−0.16 points13%†
GPT-6.1 Terra high · chat13.4%−0.16 points13%†
Claude Opus 5.5 low · chat · 1 per request12.1%−0.11 points12%†
Gemini 3.8 Flash low · chat12.9%−0.17 points13%†
GPT-6.1 Terra low · chat16.1%−0.15 points16%†
Claude Opus 5.5 low · chat14.4%−0.20 points14%†
Claude Sonnet 5.5 high · chat · 1 per request13.9%−0.21 points14%†
Claude Sonnet 5.5 high · chat21.1%−0.07 points21%†

† Leaves out the first model’s hardware cost, which we have not estimated.