Chat model

Gemma 4 E4B default reasoning · chat

default reasoning · 1 question per request

Accuracy, all 2,198 questions
57.6%55.5% to 59.6%, 95% interval; 1,265 right
Answers kept at 99% confidence
Not measureda plain answer line, no confidence
Cost, $ per 1,000 questionsper 1M questions
$0.307$307Basis and price under Cost
Latency
13.4 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
our GPU (logic's server, vLLM v0.30.0 BF16, revision google/gemma-4-E4B-it@4ecc2ba), one question per request, plain answer line CEVAP: X (no answer schema); sampling temperature 0.0; output cap 8,192
Paid as
Our own hardware
Model ID
gemma-4-e4b-it
Parameters
8B dense · 4.5B effective, as the maker states it (counted from the released files: 8.00B).language model with per-layer embeddings + vision and audio encoders; effective size as Google states it; source: counted from the checkpoint's safetensors, google/gemma-4-E4B-it@4ecc2ba, huggingface.co/google/gemma-4-E4B-it; read 9 November 2026
Hardware
1 × RTX PRO 6000 Blackwell (96 GB); BF16
Confidence from
Not measureda plain answer line, no confidence
Run
1040, 12 November 2026

Settings

Reasoning
default
Request format
1 question per request
Requests at once
32
Temperature
not recorded
Run record: harness seed, harness commit and prompt version
Harness seed
not recorded
Harness commit
6ad0eac51a4a
Prompt version
ba455bb00e5c

Results

QuestionsAccuracy95% intervalRight
All questions57.6%55.5% to 59.6%1,265 of 2,198
Reading77.6%74.2% to 80.6%509 of 656
Grammar49.0%46.5% to 51.5%756 of 1,542

Comparison group: QRSTUVWXY. Compare with another setup.

Invalid answers: 3; read from broken replies: 0; empty replies asked again: 0. Still invalid: 3 unreadable. An empty reply is asked once more; in a broken reply, any letter that is clearly readable is scored and the rest count as wrong. Invalid answers count as wrong (how).

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading68.7%90 of 131
Sentence meaningsReading87.0%134 of 154
How a writer explains and develops ideasReading77.1%54 of 70
Topic and main point of a paragraphReading87.8%86 of 98
How a paragraph fits togetherReading60.4%64 of 106
Details that support a paragraph’s main pointReading83.5%81 of 97
Word types, such as nouns and verbsGrammar57.9%81 of 140
Groups of words built around a nounGrammar55.4%36 of 65
Verbs and their formsGrammar62.6%57 of 91
Verb forms used like nouns or modifiersGrammar61.1%44 of 72
Word endings and what they doGrammar45.6%36 of 79
How words are builtGrammar31.9%22 of 69
The roles of words and phrases in a sentenceGrammar51.7%75 of 145
How verbs show who acts and who receives the actionGrammar54.1%40 of 74
Types of sentencesGrammar55.3%47 of 85
Questions combining several grammar topicsGrammar47.3%104 of 220
Sound patterns and changesGrammar19.5%15 of 77
SpellingGrammar39.8%53 of 133
PunctuationGrammar45.3%53 of 117
Problems in sentence wordingGrammar53.1%93 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
Gemma 4 E4B default reasoning · chat Open · 8.0B54.4%42.9%54.1%Queued2 of 4 OCR routes done; 173 of 763 asked via Surya OCR 2

Calibration on these questions (ECE, Brier score, AUROC): not measured, a plain answer line gives no option probabilities, so accuracy only.

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
451 of 547, first English run (889)
Accuracy
82.4%
Non-hallucination with “I don't know”
9.8%: 10 “I don't know” of 102 not right (run 890)
Answered with “I don't know”
97.8%: 535 of 547
Error rate with “I don't know”
17.2%: 92 of 535 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
0.199199from GPU time
Latency, median request, first English run
8.39 s; 32 requests at once
English run
889
Machine, English runs
RTX PRO 6000
Date
9 November 2026
Request format
1 question per request
Reasoning
default
Prompt version
ad6f4cc346df
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

With “I don't know” (run 890): Invalid answers: 2; read from broken replies: 0; empty replies asked again: 0. Still invalid: 2 unreadable. (how)

Confidence

Not measured: this setup was asked for a plain answer line, so it states no confidence and reports no option probabilities. It has no answers kept at a confidence threshold, no calibration gap, ranking score (AUROC) or Brier score, and no typical score of a right or wrong answer.

Repeats

We did not repeat this setup.

Latency

Latency, median request
13.4 s (this run)
Latency, 90th percentile
20.4 s
Requests at once
32
Throughput
135.0 questions per minute of request time
Request time
16.3 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
$0.307$307
Whole run
$0.68
Basis
Estimated from GPU time: the run's request time x the GPUs its server held x Runpod's on-demand price for an RTX PRO 6000, $2.49 per GPU-hour (read 5 November 2026). A cloud-equivalent estimate, not measured spending.
Price
$2.49 per GPU-hour, 1 GPU, for the run's request time (source, 2026-11-05)
Input tokens per question
248
Output tokens per question
811