Decision model
Karpo Decision 31B
1 question per request
- Accuracy, all 2,198 questions
- 77.8%76.0% to 79.5%, 95% interval; 1,710 right
- Answers kept at 99% confidence
- 505 of 2,1980 wrong; share right 99.2% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.105$105Basis and price under Cost
- Latency
- 0.159 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- karpo-decision server (the System One server of the card's karpo-decision package, machinepulse-ai/karpo-decision at ac40d39), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
karpo-decision- Parameters
- 31B dense, as the maker states it (counted from the released files: 31.27B).language model + vision encoder (575,743,536) + 255-way readout (1,370,880); the card says 31.3B in bf16 plus the readout; source: counted from the checkpoint's safetensors, MachinePulseAI/Karpo Decision 31B@d69cdd5, huggingface.co/MachinePulseAI/Karpo Decision 31B; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- MachinePulse's released checkpoint, a fine-tune of Gemma-4-31B-it with a 255-way readout that also reads images, run on our hardware through the System One server of the package its card ships (karpo-decision at ac40d39). Its Reading runs shared the server with the Logic & Math site's runs from 05:00 UTC on 15 November 2026, so its latency and GPU-time cost were measured under that load.
- Confidence from
- Option score
- Run
- 1156, 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
815031979442- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 77.8% | 76.0% to 79.5% | 1,710 of 2,198 |
| Reading | 93.4% | 91.3% to 95.1% | 613 of 656 |
| Grammar | 71.1% | 68.8% to 73.3% | 1,097 of 1,542 |
Comparison group: rstu. 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
- 504 of 547, first English run (1159)
- Accuracy
- 92.1%
- Non-hallucination with “I don't know”
- 0.0%: 0 “I don't know” of 42 not right (run 1161)
- Answered with “I don't know”
- 100.0%: 547 of 547
- Error rate with “I don't know”
- 7.7%: 42 of 547 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.145145from GPU time
- Latency, median request, first English run
- 0.207 s; 1 request at once
- English run
- 1159
- 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
- 505, of which 0 wrong (share right: 99.2% to 100.0%, 95% interval); 1,693 skipped, the model less than 99% sure
- Calibration gap
- 2.3 points
- AUROC, all questions
- 0.88 0.86 to 0.89
- AUROC, reading
- 0.93 0.90 to 0.96
- AUROC, grammar
- 0.84 0.82 to 0.86
- Brier score
- 0.115
- Accuracy of its surest answers
- most confident fifth: 100.0% right; most confident half: 98.0% right; most confident four fifths: 87.5% right
- Typical score of a right answer
- 95.1%
- Typical score of a wrong answer
- 54.0%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.159 s (this run)
- Latency, 90th percentile
- 0.207 s
- Requests at once
- 1
- Throughput
- 395.9 questions per minute of request time
- Request time
- 5.6 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
- $0.105$105
- Whole run
- $0.23
- 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
- 426
- Output tokens per question
- 0