Chat model
Qwen3.5-4B reasoning on · chat
reasoning on · 1 question per request
- Accuracy, all 2,198 questions
- 58.9%56.8% to 61.0%, 95% interval; 1,295 right
- Answers kept at 99% confidence
- Not measureda plain answer line, no confidence
- Cost, $ per 1,000 questionsper 1M questions
- $2.998$2,998Basis and price under Cost
- Latency
- 140.3 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 on; output cap 32,768
- Paid as
- Our own hardware
- Model ID
qwen3.5-4b- 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
- 1149, 14 November 2026
Settings
- Reasoning
- on
- 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 | 58.9% | 56.8% to 61.0% | 1,295 of 2,198 |
| Reading | 79.1% | 75.8% to 82.1% | 519 of 656 |
| Grammar | 50.3% | 47.8% to 52.8% | 776 of 1,542 |
Comparison group: PQRSTUVW. Compare with another setup.
Invalid answers: 5; read from broken replies: 0; empty replies asked again: 0. Still invalid: 1 cut off at the output limit, 4 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 1150): Invalid answers: 5; read from broken replies: 0; empty replies asked again: 0. Still invalid: 2 cut off at the output limit, 3 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
- 476 of 547, first English run (1072)
- Accuracy
- 87.0%
- Non-hallucination with “I don't know”
- 2.4%: 2 “I don't know” of 85 not right (run 1073)
- Answered with “I don't know”
- 99.6%: 545 of 547
- Error rate with “I don't know”
- 15.2%: 83 of 545 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 1.3901,390from GPU time
- Latency, median request, first English run
- 58.3 s; 32 requests at once
- English run
- 1072
- Machine, English runs
- RTX PRO 6000
- Date
- 13 November 2026
- Request format
- 1 question per request
- Reasoning
- on
- 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
- 140.3 s (this run)
- Latency, 90th percentile
- 204.2 s
- Requests at once
- 32
- Throughput
- 13.8 questions per minute of request time
- Request time
- 158.8 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
- $2.998$2,998
- Whole run
- $6.59
- 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
- 264
- Output tokens per question
- 7,397