Chat model
Qwen3.5-4B reasoning off · chat
reasoning off · 1 question per request
- Accuracy, all 2,198 questions
- 46.2%44.1% to 48.3%, 95% interval; 1,015 right
- Answers kept at 99% confidence
- Not measureda plain answer line, no confidence
- Cost, $ per 1,000 questionsper 1M questions
- $0.894$894Basis and price under Cost
- Latency
- 32.2 smedian request, 1 question per request
How we called it
- Kind
- Chat model
- Access
- our GPU (logic's server, vLLM v0.30.0 BF16, revision Qwen/Qwen3.5-4B@851bf6e), one question per request, plain answer line CEVAP: X (no answer schema); sampling min_p 0.0, presence_penalty 1.5, repetition_penalty 1.0, seed 0, temperature 1.0, top_k 20, top_p 0.95, thinking off; output cap 32,768
- Paid as
- Our own hardware
- Model ID
qwen3.5-4b-nothink- Parameters
- 4B dense, as the maker states it (counted from the released files: 4.66B).language model (tied embeddings) + multi-token prediction layer + 0.33B vision encoder; source: counted from the checkpoint's safetensors, Qwen/Qwen3.5-4B@851bf6e, huggingface.co/Qwen/Qwen3.5-4B; read 9 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Confidence from
- Not measureda plain answer line, no confidence
- Run
- 1146, 14 November 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 32
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
c47b12d3647b- Prompt version
ba455bb00e5c
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 46.2% | 44.1% to 48.3% | 1,015 of 2,198 |
| Reading | 68.4% | 64.8% to 71.9% | 449 of 656 |
| Grammar | 36.7% | 34.3% to 39.1% | 566 of 1,542 |
Comparison group: θ. Compare with another setup.
Invalid answers: 21; read from broken replies: 0; empty replies asked again: 0. Still invalid: 12 cut off at the output limit, 9 unreadable. 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).
With “Bilmiyorum” (run 1147): Invalid answers: 18; read from broken replies: 0; empty replies asked again: 0. Still invalid: 8 cut off at the output limit, 10 unreadable. (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
- 440 of 547, first English run (1143)
- Accuracy
- 80.4%
- Non-hallucination with “I don't know”
- 3.6%: 4 “I don't know” of 110 not right (run 1144)
- Answered with “I don't know”
- 99.3%: 543 of 547
- Error rate with “I don't know”
- 19.5%: 106 of 543 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.309309from GPU time
- Latency, median request, first English run
- 9.74 s; 32 requests at once
- English run
- 1143
- Machine, English runs
- RTX PRO 6000
- Date
- 14 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
ad6f4cc346df- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
Not measured: this setup was asked for a plain answer line, so it states no confidence and reports no option probabilities. It has no answers kept at a confidence threshold, no calibration gap, ranking score (AUROC) or Brier score, and no typical score of a right or wrong answer.
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 32.2 s (this run)
- Latency, 90th percentile
- 74.8 s
- Requests at once
- 32
- Throughput
- 46.4 questions per minute of request time
- Request time
- 47.3 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
- $0.894$894
- Whole run
- $1.96
- 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
- 266
- Output tokens per question
- 2,186