Chat model
DeepSeek-V4.1-Flash reasoning on · chat · 1017
reasoning on · 16 questions per request
- Accuracy, all 2,198 questions
- 91.8%90.5% to 92.8%, 95% interval; 2,017 right
- Answers kept at 99% confidence
- 2,164 of 2,198149 wrong; share right 92.0% to 94.1%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $1.278$1,278Basis and price under Cost
- Latency
- 486.9 smedian request, 16 questions per request
How we called it
- Kind
- Chat model
- Access
- SGLang, self-hosted, published mixed precision; 1017 weights (revision 1bae0e1)
- Paid as
- Our own hardware
- Model ID
deepseek-v4.1-flash-1017- Parameters
- Mixture of experts: 552B total, about 16B active per step. Size class uses the total (552B).V4.1 Flash card (first release): 552B backbone + 196B conditional memory; 8B active in prefill, 16B in decode; the 1017 card states no size; pages checked: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017, huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash; read 9 November 2026
- Hardware
- 4 × RTX PRO 6000 Blackwell (96 GB); FP4 experts, FP8 and BF16 layers
- Confidence from
- Answer letter probability
- Run
- 585, 5 November 2026
Settings
- Reasoning
- on
- Request format
- 16 questions per request
- Requests at once
- 8
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
d4268ef9b103- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 91.8% | 90.5% to 92.8% | 2,017 of 2,198 |
| Reading | 94.1% | 92.0% to 95.6% | 617 of 656 |
| Grammar | 90.8% | 89.2% to 92.1% | 1,400 of 1,542 |
Comparison group: jkl. Compare with another setup.
Invalid answers: 32; read from broken replies: 0; empty replies asked again: 0. Still invalid: 16 empty and not yet asked again, 16 cut off at the output limit. An empty reply is asked once more; in a broken reply, any letter that is clearly readable is scored and the rest count as wrong. Invalid answers count as wrong (how).
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
- 506 of 547, first English run (808)
- Accuracy
- 92.5%
- Non-hallucination with “I don't know”
- 7.1%: 3 “I don't know” of 42 not right (run 809)
- Answered with “I don't know”
- 99.5%: 544 of 547
- Error rate with “I don't know”
- 7.2%: 39 of 544 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.421421at the maker's list price
- Latency, median request, first English run
- 189.4 s; 8 requests at once
- English run
- 808
- Machine, English runs
- RTX PRO 6000
- Date
- 5 November 2026
- Request format
- 16 questions per request
- Reasoning
- on
- Prompt version
5c3a4849752e- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 2,164, of which 149 wrong (share right: 92.0% to 94.1%, 95% interval); 34 skipped: 2 with the model less than 99% sure and 32 without a valid answer
- Calibration gap
- 6.9 points
- AUROC, all questions
- 0.61 0.57 to 0.66
- AUROC, reading
- 0.61 0.51 to 0.70
- AUROC, grammar
- 0.62 0.56 to 0.68
- Brier score
- 0.069
- Accuracy of its surest answers
- most confident fifth: 95.4% right; most confident half: 95.5% right; most confident four fifths: 94.2% right
- Typical score of a right answer
- 100.0%
- Typical score of a wrong answer
- 100.0%
Repeats
Run 5 times on a sample of 200 questions.
- Answers changed in 5 runs
- 15 of 200
- Same wrong answer in all 5
- 1 of 200
- Wrong in at least one run
- 16 of 200
- Accuracy in each run
- 96.5%, 94.5%, 97.0%, 95.0%, 96.0%
Latency
- Latency, median request
- 486.9 s (this run)
- Latency, 90th percentile
- 1129.4 s
- Requests at once
- 8
- Throughput
- 12.9 questions per minute of request time
- Request time
- 170.9 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
- $1.278$1,278
- Whole run
- $2.81
- Basis
- Estimated at DeepSeek's list price for the same model; our hardware cost is not measured.
- Price
- $0.30 input, $0.006 cached input, $1.20 output per 1M tokens (source, 2026-11-05)
- Input tokens per question
- 305
- Output tokens per question
- 989