Decision model
vLLM-SR d3
1 question per request
67.8% of the 2,198 questions right (65.8% to 69.7%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- vLLM System One endpoint (nightly), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
vllm-sr/d3- Parameters
- 26.1B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- The vLLM Semantic Router's released Decision 3.0 checkpoint (vllm-sr/d3 at 20f37b9), fine-tuned from Qwen3.8-27B (Qwen3.5 architecture), served on our hardware through vLLM's System One endpoint by a nightly build (0.32.1rc1.dev613+g8977b13a1) with --max-num-seqs 256: with the default, vLLM's startup memory check did not fit the GPU. One request at a time, as for the other setups.
- Confidence from
- Option score
- Run
- 259, 27 October 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
- Harness seed
- 0
- Harness commit
4a94c95fff4b- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 67.8% | 65.8% to 69.7% | 1,490 of 2,198 |
| Reading | 89.2% | 86.6% to 91.3% | 585 of 656 |
| Grammar | 58.7% | 56.2% to 61.1% | 905 of 1,542 |
Comparison group: pqr. Compare with another setup.
Accuracy by topic
Page images (preliminary)
Preliminary: these numbers are not final and may change. This setup is condition A of the Page images tab, which asks the 763 questions with underlined or numbered words as text (A), as a page image (B), as both (C) and through OCR models (D); the share answered right, with its 95% interval.
| Model | A · text onlyaccuracy, higher is better | B · image onlyaccuracy, higher is better | C · text plus imageaccuracy, higher is better | Best OCR routeaccuracy, higher is better |
|---|---|---|---|---|
| vLLM-SR d31 question per request | 63.2%59.7% to 66.5% | 48.5%45.0% to 52.0% | 64.9%61.4% to 68.2% | 60.2%via Surya OCR 256.6% to 63.6% |
Confidence
- Answers kept at 99% or more
- 250, of which 0 wrong (share right: 98.5% to 100.0%, 95% interval); 1,948 skipped, the model less than 99% sure
- Calibration gap
- 2.6 points
- AUROC, all questions
- 0.84 0.82 to 0.85
- AUROC, reading
- 0.93 0.91 to 0.95
- AUROC, grammar
- 0.77 0.75 to 0.79
- Brier score
- 0.153
- Accuracy of its surest answers
- most confident fifth: 100.0% right; most confident half: 91.1% right; most confident four fifths: 76.1% right
- Typical score of a right answer
- 85.8%
- Typical score of a wrong answer
- 46.5%
Repeats
Run 5 times on all questions.
- Answers changed in 5 runs
- 0 of 2,198
- Same wrong answer in all 5
- 708 of 2,198
- Wrong in at least one run
- 708 of 2,198
- Accuracy in each run
- 67.8%, 67.8%, 67.8%, 67.8%, 67.8%
Latency
- Latency, median request
- 0.092 s (this run)
- Latency, 90th percentile
- 0.144 s
- Requests at once
- 1
- Throughput
- 591.1 questions per minute of request time
- Request time
- 3.7 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
- $ per 1,000 questions
- Not measured
- Basis
- No dollar estimate; see hardware and time
- Input tokens per question
- 393
- Output tokens per question
- 0
Going first in front of an LLM
Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.
| LLM behind it | Share taken | Accuracy change | Saving |
|---|---|---|---|
| Gemini 3.8 Flashhigh reasoning · 16 questions per request | 21.9% | −0.09 points | 22%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 22.9% | −0.09 points | 23%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 24.3% | −0.11 points | 24%† |
| GPT-6 Astralow reasoning · 16 questions per request | 23.4% | −0.10 points | 23%† |
| GPT-6 Astralow reasoning · 1 question per request | 21.9% | −0.09 points | 22%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 23.0% | −0.12 points | 23%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 25.6% | −0.14 points | 26%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 23.4% | −0.10 points | 23%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 24.5% | −0.18 points | 25%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 24.2% | −0.12 points | 24%† |
| Claude Opus 5.5low reasoning · 1 question per request | 26.7% | −0.16 points | 27%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 25.0% | −0.20 points | 25%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 32.3% | −0.16 points | 32%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 30.8% | −0.20 points | 31%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 31.6% | −0.17 points | 32%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 50.2% | −0.07 points | 50%† |
† Leaves out the first model’s hardware cost, which we have not estimated.