Decision model
OneJev 27B
1 question per request
59.6% of the 2,198 questions right (57.5% to 61.6%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- OneJev server (qev, OmniJev's release code), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
OmniJev/OneJev-27B- Parameters
- 27.4B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- OmniJev's released OneJev checkpoint (OmniJev/OneJev-27B at a5e5f55), a full fine-tune of Qwen3.8-27B that also reads images, served on our hardware through the server its package ships (qev 0.1.4, qev serve) with its defaults. One request at a time, as for the other setups. The five runs are identical because decoding is deterministic: each run sent all 2,198 questions again as separate requests.
- Confidence from
- Option score
- Run
- 264, 27 October 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
- Harness seed
- 0
- Harness commit
9162663222cc- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 59.6% | 57.5% to 61.6% | 1,310 of 2,198 |
| Reading | 85.2% | 82.3% to 87.7% | 559 of 656 |
| Grammar | 48.7% | 46.2% to 51.2% | 751 of 1,542 |
Comparison group: vwxyz. Compare with another setup.
Accuracy by topic
Page images (preliminary)
Preliminary: these numbers are not final and may change. This setup is condition A of the Page images 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, with its 95% interval.
| Model | A · text onlyaccuracy, higher is better | B · image onlyaccuracy, higher is better | C · text plus imageaccuracy, higher is better | Best OCR routeaccuracy, higher is better |
|---|---|---|---|---|
| OneJev 27B1 question per request | 54.0%50.4% to 57.5% | 41.3%37.8% to 44.8% | 58.7%55.2% to 62.2% | 50.6%via Surya OCR 247.0% to 54.1% |
Confidence
- Answers kept at 99% or more
- 2, of which 0 wrong (share right: 34.2% to 100.0%, 95% interval); 2,196 skipped, the model less than 99% sure
- Calibration gap
- 5.5 points
- AUROC, all questions
- 0.82 0.80 to 0.84
- AUROC, reading
- 0.91 0.88 to 0.94
- AUROC, grammar
- 0.74 0.71 to 0.76
- Brier score
- 0.170
- Accuracy of its surest answers
- most confident fifth: 98.9% right; most confident half: 83.8% right; most confident four fifths: 68.1% right
- Typical score of a right answer
- 69.5%
- Typical score of a wrong answer
- 34.3%
Repeats
Run 5 times on all questions.
- Answers changed in 5 runs
- 0 of 2,198
- Same wrong answer in all 5
- 888 of 2,198
- Wrong in at least one run
- 888 of 2,198
- Accuracy in each run
- 59.6%, 59.6%, 59.6%, 59.6%, 59.6%
Latency
- Latency, median request
- 0.125 s (this run)
- Latency, 90th percentile
- 0.189 s
- Requests at once
- 1
- Throughput
- 418.8 questions per minute of request time
- Request time
- 5.2 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
- Not measured
- Basis
- No dollar estimate; see hardware and time
- Input tokens per question
- 396
- Output tokens per question
- 1
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 it | Share taken | Accuracy change | Saving |
|---|---|---|---|
| Gemini 3.8 Flashhigh reasoning · 16 questions per request | 17.1% | −0.10 points | 17%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 17.3% | −0.12 points | 17%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 18.8% | −0.15 points | 19%† |
| GPT-6 Astralow reasoning · 16 questions per request | 17.3% | −0.12 points | 17%† |
| GPT-6 Astralow reasoning · 1 question per request | 17.1% | −0.10 points | 17%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 17.7% | −0.13 points | 18%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 17.5% | −0.13 points | 18%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 17.4% | −0.12 points | 17%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 17.2% | −0.10 points | 17%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 17.3% | −0.12 points | 17%† |
| Claude Opus 5.5low reasoning · 1 question per request | 19.2% | −0.21 points | 19%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 18.3% | −0.14 points | 18%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 20.5% | −0.14 points | 21%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 19.0% | −0.19 points | 19%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 22.0% | −0.24 points | 22%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 33.7% | −0.18 points | 34%† |
† Leaves out the first model’s hardware cost, which we have not estimated.