Decision model
vLLM-SR d3-ultra
transformers, rev 52ba9e7 · 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
- d3 package server (transformers, the release code), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
vllm-sr/d3-ultra- Variant
- transformers, rev 52ba9e7
- Parameters
- 26.1B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- The vLLM Semantic Router's released d3-ultra checkpoint (vllm-sr/d3-ultra at 52ba9e7), fine-tuned from Qwen3.8-27B (Qwen3.5 architecture), served on our hardware through the server its repository ships (d3_server.py, the package's transformers runtime as in the model card's Quickstart; transformers 5.17.0). vLLM cannot serve it yet: it rejects the checkpoint's attention mode (prefix_bidir). One request at a time, as for the other setups.
- Confidence from
- Option score
- Run
- 269, 27 October 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
- Harness seed
- 0
- Harness commit
0fbd97b0899d- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 67.8% | 65.8% to 69.7% | 1,490 of 2,198 |
| Reading | 89.8% | 87.2% to 91.9% | 589 of 656 |
| Grammar | 58.4% | 56.0% to 60.9% | 901 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 d3-ultratransformers, rev 52ba9e7 · 1 question per request | 63.0%59.6% to 66.4% | 45.5%42.0% to 49.0% | 64.5%61.0% to 67.8% | 58.7%via Surya OCR 255.2% to 62.2% |
Confidence
- Answers kept at 99% or more
- 384, of which 1 wrong (share right: 98.5% to 100.0%, 95% interval); 1,814 skipped, the model less than 99% sure
- Calibration gap
- 12.5 points
- AUROC, all questions
- 0.83 0.82 to 0.85
- AUROC, reading
- 0.92 0.89 to 0.94
- AUROC, grammar
- 0.77 0.74 to 0.79
- Brier score
- 0.177
- Accuracy of its surest answers
- most confident fifth: 99.8% right; most confident half: 90.8% right; most confident four fifths: 75.9% right
- Typical score of a right answer
- 96.1%
- Typical score of a wrong answer
- 63.9%
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
- 0.333 s (this run)
- Latency, 90th percentile
- 0.440 s
- Requests at once
- 1
- Throughput
- 192.3 questions per minute of request time
- Request time
- 11.4 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
- 513
- 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 | 22.7% | −0.10 points | 23%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 24.4% | −0.10 points | 24%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 26.3% | −0.14 points | 26%† |
| GPT-6 Astralow reasoning · 16 questions per request | 25.5% | −0.15 points | 26%† |
| GPT-6 Astralow reasoning · 1 question per request | 23.0% | −0.12 points | 23%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 25.1% | −0.15 points | 25%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 26.5% | −0.14 points | 26%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 25.6% | −0.16 points | 26%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 28.0% | −0.20 points | 28%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 25.9% | −0.18 points | 26%† |
| Claude Opus 5.5low reasoning · 1 question per request | 26.5% | −0.13 points | 27%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 26.4% | −0.22 points | 26%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 32.6% | −0.18 points | 33%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 33.2% | −0.19 points | 33%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 29.4% | −0.26 points | 29%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 52.6% | −0.15 points | 53%† |
† Leaves out the first model’s hardware cost, which we have not estimated.