Decision model
Logits Sigma-1 Duo
1 question per request
- Accuracy, all 2,198 questions
- 75.0%73.2% to 76.8%, 95% interval; 1,649 right
- Answers kept at 99% confidence
- 42 of 2,1980 wrong; share right 91.6% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.181$181Basis and price under Cost
- Latency
- 0.250 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Decision Index engine (compactor's di_serve.py around the engine its card ships), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
Logits-Sigma-1-Duo- Parameters
- 57.36B dense, as the maker states it.both members: KnowLine-31B-Gen1 with Logits' LoRA merged, and d3-ultra; source: the Jev Decision Index's served_params for logitshq/Logits-Sigma-1-Duo@a0c1b6e8, huggingface.co/logitshq/Logits-Sigma-1-Duo; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); FP8
- Note
- Logits' released Sigma-1 Duo (logitshq/Logits-Sigma-1-Duo at a0c1b6e, the weights and code of d9641cd), two decision models on one GPU whose per-option probabilities it averages 50/50: KnowLine-31B-Gen1 with Logits' LoRA merged, and vLLM-SR d3-ultra. It ran on our hardware in FP8 through its card's own engine and image (plus gcc, libc6-dev and python3-dev, which its Triton kernels need) behind compactor's di_serve.py, text only; the engine took its own Triton FP8 kernel on this GPU.
- Confidence from
- Option score
- Run
- 1221, 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
7b06c24e4bf8- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 75.0% | 73.2% to 76.8% | 1,649 of 2,198 |
| Reading | 91.9% | 89.6% to 93.8% | 603 of 656 |
| Grammar | 67.8% | 65.5% to 70.1% | 1,046 of 1,542 |
Comparison group: suvwx. 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
- 498 of 547, first English run (1229)
- Accuracy
- 91.0%
- Non-hallucination with “I don't know”
- 4.3%: 2 “I don't know” of 46 not right (run 1231)
- Answered with “I don't know”
- 99.6%: 545 of 547
- Error rate with “I don't know”
- 8.1%: 44 of 545 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.191191from GPU time
- Latency, median request, first English run
- 0.263 s; 1 request at once
- English run
- 1229
- 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
- 42, of which 0 wrong (share right: 91.6% to 100.0%, 95% interval); 2,156 skipped, the model less than 99% sure
- Calibration gap
- 5.7 points
- AUROC, all questions
- 0.85 0.84 to 0.87
- AUROC, reading
- 0.93 0.91 to 0.95
- AUROC, grammar
- 0.81 0.78 to 0.83
- Brier score
- 0.133
- Accuracy of its surest answers
- most confident fifth: 100.0% right; most confident half: 96.6% right; most confident four fifths: 83.4% right
- Typical score of a right answer
- 84.1%
- Typical score of a wrong answer
- 48.0%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.250 s (this run)
- Latency, 90th percentile
- 0.324 s
- Requests at once
- 1
- Throughput
- 229.3 questions per minute of request time
- Request time
- 9.6 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
- $0.181$181
- Whole run
- $0.40
- 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
- 0
- Output tokens per question
- 0