Decision model
Sieve 31B v3
1 question per request
- Accuracy, all 2,198 questions
- 74.2%72.3% to 75.9%, 95% interval; 1,630 right
- Answers kept at 99% confidence
- 329 of 2,1981 wrong; share right 98.3% to 99.9%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- No list priceBasis and price under Cost
- Latency
- 1.00 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Sieve server (sieve-serve, the maker's release code), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
sthanika-ai/Sieve-31B-v3- Parameters
- 31B dense, as the maker states it (counted from the released files: 31.27B).language model (tied embeddings) + 0.58B vision encoder (not read by its server); source: counted from the checkpoint's safetensors, sthanika-ai/Sieve-31B-v3@8fcee85, huggingface.co/sthanika-ai/Sieve-31B-v3; read 10 November 2026
- Hardware
- 1 × DGX Spark, GB10 (128 GB unified memory); BF16
- Note
- sthanika-ai's released Sieve-31B-v3 (sthanika-ai/Sieve-31B-v3 at 8fcee85), a full fine-tune of Gemma-4-31B-it with bidirectional attention in its global layers, read in two option orders. It ran on our hardware (one DGX Spark), one request at a time, through the server its card names (sieve-serve from the Sieve repository's main at 1e64657, on torch 2.14 where the card tested torch 2.8), text only. Probabilities rounded to 4 decimals by its own code; they sum to 1 only within that rounding. Its server reports no token counts. Its card lists English only.
- Confidence from
- Option score
- Run
- 1241, 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
28d1c0e8e51a- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 74.2% | 72.3% to 75.9% | 1,630 of 2,198 |
| Reading | 91.8% | 89.4% to 93.6% | 602 of 656 |
| Grammar | 66.7% | 64.3% to 69.0% | 1,028 of 1,542 |
Comparison group: vwxy. 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
- 496 of 547, first English run (1064)
- Accuracy
- 90.7%
- Non-hallucination with “I don't know”
- 2.2%: 1 “I don't know” of 46 not right (run 1066)
- Answered with “I don't know”
- 99.8%: 546 of 547
- Error rate with “I don't know”
- 8.2%: 45 of 546 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- No list price
- Latency, median request, first English run
- 1.06 s; 1 request at once
- English run
- 1064
- Machine, English runs
- DGX Spark
- Date
- 12 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 329, of which 1 wrong (share right: 98.3% to 99.9%, 95% interval); 1,869 skipped, the model less than 99% sure
- Calibration gap
- 2.9 points
- AUROC, all questions
- 0.87 0.86 to 0.89
- AUROC, reading
- 0.94 0.91 to 0.96
- AUROC, grammar
- 0.83 0.81 to 0.85
- Brier score
- 0.126
- Accuracy of its surest answers
- most confident fifth: 99.8% right; most confident half: 96.5% right; most confident four fifths: 84.5% right
- Typical score of a right answer
- 92.4%
- Typical score of a wrong answer
- 49.2%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 1.00 s (this run)
- Latency, 90th percentile
- 1.37 s
- Requests at once
- 1
- Throughput
- 56.4 questions per minute of request time
- Request time
- 39.0 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
- No list price
- Basis
- No dollar estimate; see hardware and time
- Input tokens per question
- 0
- Output tokens per question
- 0