Chat model

Qwen3.8-27B

vote share · reasoning off · 1 question per request

54.3% of the 2,198 questions right (52.2% to 56.4%, 95% interval). Back to the leaderboard.

How we called it

Kind: Chat model
Access: vLLM, self-hosted, BF16
Paid as: Our own hardware
Model ID: qwen38-27b-bf16
Variant: vote share
Parameters: 27.8B
Hardware: 1 × RTX PRO 6000 Blackwell (96 GB); BF16
Note: Open-Jev's base model
Confidence from: Agreement across ten answers
Run: 83, 4 October 2026

Settings

Reasoning: off
Request format: 1 question per request
Requests at once: 1
Temperature: not recorded
Harness seed: 0
Harness commit: not recorded
Prompt version: not recorded

Results

Questions	Accuracy	95% interval	Right
All questions	54.3%	52.2% to 56.4%	1,194 of 2,198
Reading	84.0%	81.0% to 86.6%	551 of 656
Grammar	41.7%	39.3% to 44.2%	643 of 1,542

Comparison group: BC. Compare with another setup.

Accuracy by topic

Topic	Part	Accuracy	Right
Word and phrase meanings	Reading	80.9%	106 of 131
Sentence meanings	Reading	86.4%	133 of 154
How a writer explains and develops ideas	Reading	74.3%	52 of 70
Topic and main point of a paragraph	Reading	92.9%	91 of 98
How a paragraph fits together	Reading	77.4%	82 of 106
Details that support a paragraph’s main point	Reading	89.7%	87 of 97
Word types, such as nouns and verbs	Grammar	42.1%	59 of 140
Groups of words built around a noun	Grammar	33.8%	22 of 65
Verbs and their forms	Grammar	46.2%	42 of 91
Verb forms used like nouns or modifiers	Grammar	37.5%	27 of 72
Word endings and what they do	Grammar	41.8%	33 of 79
How words are built	Grammar	36.2%	25 of 69
The roles of words and phrases in a sentence	Grammar	40.7%	59 of 145
How verbs show who acts and who receives the action	Grammar	44.6%	33 of 74
Types of sentences	Grammar	43.5%	37 of 85
Questions combining several grammar topics	Grammar	23.6%	52 of 220
Sound patterns and changes	Grammar	31.2%	24 of 77
Spelling	Grammar	50.4%	67 of 133
Punctuation	Grammar	50.4%	59 of 117
Problems in sentence wording	Grammar	59.4%	104 of 175

Confidence

Answers kept at 99% or more: 875, of which 129 wrong (share right: 82.8% to 87.5%, 95% interval); 1,323 skipped, the model less than 99% sure
Calibration gap: 22.5 points
AUROC, all questions: 0.78 0.76 to 0.80
AUROC, reading: 0.80 0.75 to 0.85
AUROC, grammar: 0.69 0.67 to 0.72
Brier score: 0.241
Accuracy of its surest answers: most confident fifth: 85.3% right; most confident half: 76.9% right; most confident four fifths: 61.3% right
Typical score of a right answer: 100.0%
Typical score of a wrong answer: 60.0%

Repeats

We did not repeat this setup.

Latency

Latency, median request: 1.62 s (this run)
Latency, 90th percentile: 1.84 s
Requests at once: 1
Throughput: 36.4 questions per minute of request time
Request time: 60.3 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

$ per 1,000 questions: $0.83
Whole run: $1.81
Basis: Estimated at Alibaba's list price for the same model; our hardware cost is not measured.
Price: $0.50 input, $0.10 cached input, $3 output per 1M tokens (source, 2026-10-18)
Input tokens per question: 336
Output tokens per question: 219