Decision model
Turion Jev 31B
1 question per request
- Accuracy, all 2,198 questions
- 80.1%78.4% to 81.7%, 95% interval; 1,761 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
- No list priceBasis and price under Cost
- Latency
- 0.405 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Turion Jev server (turion_jev.serve on vLLM, the maker's release code), self-hosted, FP8 at load
- Paid as
- Our own hardware
- Model ID
turion-ai/turion-jev-31b- 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, turion-ai/turion-jev-31b@866c0fc, huggingface.co/turion-ai/turion-jev-31b; read 10 November 2026
- Hardware
- 1 × DGX Spark, GB10 (128 GB unified memory); FP8 at load
- Note
- Turion AI's released Turion Jev 31B (turion-ai/turion-jev-31b at 866c0fc), a merged LoRA fine-tune of Gemma-4-31B-it. It ran on our hardware (one DGX Spark), one request at a time, through the server its package ships (python -m turion_jev.serve on vLLM), with the BF16 weights quantised to FP8 at load and --no-recheck (no second reading of unsure questions), both as in the card's own benchmark run, text only. Probabilities rounded to 6 decimals by its own code; they sum to 1 only within that rounding.
- Confidence from
- Option score
- Run
- 1273, 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
88e48d7a1055- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 80.1% | 78.4% to 81.7% | 1,761 of 2,198 |
| Reading | 92.7% | 90.4% to 94.4% | 608 of 656 |
| Grammar | 74.8% | 72.5% to 76.9% | 1,153 of 1,542 |
Comparison group: opqr. 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 (1075)
- Accuracy
- 90.7%
- Non-hallucination with “I don't know”
- 8.5%: 4 “I don't know” of 47 not right (run 1077)
- Answered with “I don't know”
- 99.3%: 543 of 547
- Error rate with “I don't know”
- 7.9%: 43 of 543 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- No list price
- Latency, median request, first English run
- 0.445 s; 1 request at once
- English run
- 1075
- Machine, English runs
- DGX Spark
- Date
- 13 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
- 11.1 points
- AUROC, all questions
- 0.87 0.86 to 0.89
- AUROC, reading
- 0.92 0.88 to 0.95
- AUROC, grammar
- 0.84 0.82 to 0.86
- Brier score
- 0.126
- Accuracy of its surest answers
- most confident fifth: 99.8% right; most confident half: 97.7% right; most confident four fifths: 89.8% right
- Typical score of a right answer
- 79.6%
- Typical score of a wrong answer
- 41.2%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.405 s (this run)
- Latency, 90th percentile
- 0.566 s
- Requests at once
- 1
- Throughput
- 140.1 questions per minute of request time
- Request time
- 15.7 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
- 766
- Output tokens per question
- 0