Decision model
vera-medium-omni
1 question per request
- Accuracy, all 2,198 questions
- 49.1%47.0% to 51.2%, 95% interval; 1,079 right
- Answers kept at 99% confidence
- 4 of 2,1980 wrong; share right 51.0% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.021$20.72Basis and price under Cost
- Latency
- 0.029 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- vera server (the card's own vera-serve from vera-s1 0.1.8, PyTorch engine), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
vera-medium-omni- Parameters
- 4.2B dense, as the maker states it (counted from the released files: 4.21B).the whole checkpoint, its vision tower included (text only here); source: counted from the checkpoint's safetensors, scar-ai/vera-medium-omni@1f23a6c, huggingface.co/scar-ai/vera-medium-omni; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- scar-ai's released vera-medium-omni (scar-ai/vera-medium-omni at 1f23a6c), a decision model on Qwen3.5-4B (about 4.2B parameters for text, plus the stock 0.33B vision tower, unused here) that reads each question in one forward pass. It ran on our hardware through its card's own server, text only; that server rounds each option's probability to 4 decimals (vera/schema.py line 110), so they sum to one only within rounding.
- Confidence from
- Option score
- Run
- 1264, 15 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
c6e3cdddbbc0- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 49.1% | 47.0% to 51.2% | 1,079 of 2,198 |
| Reading | 75.5% | 72.0% to 78.6% | 495 of 656 |
| Grammar | 37.9% | 35.5% to 40.3% | 584 of 1,542 |
Comparison group: γδζηθ. Compare with another setup.
Accuracy by topic
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
- 457 of 547, first English run (1267)
- Accuracy
- 83.5%
- Non-hallucination with “I don't know”
- 7.4%: 7 “I don't know” of 94 not right (run 1269)
- Answered with “I don't know”
- 98.7%: 540 of 547
- Error rate with “I don't know”
- 16.1%: 87 of 540 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.03736.63from GPU time
- Latency, median request, first English run
- 0.054 s; 1 request at once
- English run
- 1267
- Machine, English runs
- RTX PRO 6000
- Date
- 15 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 4, of which 0 wrong (share right: 51.0% to 100.0%, 95% interval); 2,194 skipped, the model less than 99% sure
- Calibration gap
- 3.7 points
- AUROC, all questions
- 0.75 0.72 to 0.76
- AUROC, reading
- 0.82 0.78 to 0.85
- AUROC, grammar
- 0.65 0.62 to 0.68
- Brier score
- 0.203
- Accuracy of its surest answers
- most confident fifth: 84.8% right; most confident half: 66.9% right; most confident four fifths: 55.2% right
- Typical score of a right answer
- 58.8%
- Typical score of a wrong answer
- 39.7%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.029 s (this run)
- Latency, 90th percentile
- 0.033 s
- Requests at once
- 1
- Throughput
- 2003.2 questions per minute of request time
- Request time
- 1.1 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
- $0.021$20.72
- Whole run
- $0.046
- 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
- 393
- Output tokens per question
- 34