Decision model
Dexy-Decision-9B-V2
1 question per request
- Accuracy, all 2,198 questions
- 51.6%49.5% to 53.7%, 95% interval; 1,134 right
- Answers kept at 99% confidence
- 138 of 2,1981 wrong; share right 96.0% to 99.9%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.023$23.03Basis and price under Cost
- Latency
- 0.031 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Dexy server (the card's Transformers d3_server.py, in its requirements.txt pins: torch 2.14.0, transformers 5.19.0), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
Dexy-Decision-9B-V2- Parameters
- 9B dense, as the maker states it (counted from the released files: 8.39B).the whole checkpoint with its 0.456B vision encoder (text only here); its card says 8.394B; source: counted from the checkpoint's safetensors, Patrick-18/Dexy-Decision-9B-V2@4f434e2, huggingface.co/Patrick-18/Dexy-Decision-9B-V2; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Patrick-18's released Dexy-Decision-9B-V2 (Patrick-18/Dexy-Decision-9B-V2 at 4f434e2), a fine-tune of vLLM-SR's d3-flash (8.394B parameters with its 0.456B vision encoder, as its card says). It ran on our hardware through its card's own server, text only.
- Confidence from
- Option score
- Run
- 1255, 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 | 51.6% | 49.5% to 53.7% | 1,134 of 2,198 |
| Reading | 79.6% | 76.3% to 82.5% | 522 of 656 |
| Grammar | 39.7% | 37.3% to 42.2% | 612 of 1,542 |
Comparison group: αβγδζ. 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 (1258)
- Accuracy
- 83.5%
- Non-hallucination with “I don't know”
- 1.1%: 1 “I don't know” of 91 not right (run 1260)
- Answered with “I don't know”
- 99.8%: 546 of 547
- Error rate with “I don't know”
- 16.5%: 90 of 546 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.02525.41from GPU time
- Latency, median request, first English run
- 0.036 s; 1 request at once
- English run
- 1258
- 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
- 138, of which 1 wrong (share right: 96.0% to 99.9%, 95% interval); 2,060 skipped, the model less than 99% sure
- Calibration gap
- 9.1 points
- AUROC, all questions
- 0.77 0.75 to 0.79
- AUROC, reading
- 0.83 0.80 to 0.87
- AUROC, grammar
- 0.67 0.64 to 0.70
- Brier score
- 0.199
- Accuracy of its surest answers
- most confident fifth: 93.2% right; most confident half: 71.4% right; most confident four fifths: 57.6% right
- Typical score of a right answer
- 76.0%
- Typical score of a wrong answer
- 44.7%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.031 s (this run)
- Latency, 90th percentile
- 0.040 s
- Requests at once
- 1
- Throughput
- 1802.4 questions per minute of request time
- Request time
- 1.2 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
- Cost per 1,000 questionsper 1M questions
- $0.023$23.03
- Whole run
- $0.051
- 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
- 0