d3:mini
Decision 3.0 (d3) · vLLM Semantic Router · 4B · ONNX Runtime · Apache-2.0
- Score
- 68.0
- 95 % interval 66.1–69.8
- Rank
- 2
- of 66 scored models
- Public benchmark
- 78.8
- macro accuracy, %
- typed-decisions
- 73.7
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
61.4
Classification
MASSIVE en
87.9
MASSIVE de
86.1
Civil Comments
86.7
Aegis 2
60.8
Understanding
MultiNLI
87.5
VitaminC
74.0
PAWS
78.4
Knowledge
BoolQ
78.0
SQuAD 2
48.5
PubMedQA
57.4
Judging
HelpSteer 2
26.7
SummEval rel.
39.6
SummEval cons.
79.2
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 78.8(77.5–80.1) | 84.3 | 85.2 | 58.8 | 0.030 | 0.281 | 0 / 3,880 |
| typed-decisions | 73.7(71.5–75.8) | 75.2 | 81.3 | 66.8 | 0.047 | 0.374 | 0 / 400 requests |
Accuracy in %, with its 95 % bootstrap interval. A rejected request (the model can’t take the question, or the input is too long for it) counts as wrong.
Speed and memory
| Device | p50 | p90 | Load | VRAM | RAM peak | Precision |
|---|---|---|---|---|---|---|
| RTX 5090 CUDA | 202 ms | 204 ms | 17.7 s | 10.8 GB | 2.0 GB | F32 |
| Threadripper 3960X CPU | 9.59 s | 9.64 s | 21.2 s | – | 9.2 GB | F32 |
Five-question requests over HTTP, one at a time, each a different short message. VRAM: the rise of the GPU’s used memory while the model is loaded. RAM peak: the runner process’s peak resident memory.
Parity with the authors’ code
| Device | Same decision | Max logit diff. | Result | Record |
|---|---|---|---|---|
| RTX 5090 CPU | 696/696 | 3.1e-5 | pass | results/runs/2026-10-24-parity-rtx5090-cpu.json |
| RTX 5090 CUDA | 696/696 | 2.5e-5 | pass | results/runs/2026-10-24-parity-rtx5090-cuda.json |
Reproduce
The tools are in convert/ollaya_convert and need only Python’s standard library. Run them against a local Ollaya 0.13.0 with the same model digest.
ollaya pull d3:mini
# typed-decisions; export the test split once with
# uv run python -m ollaya_convert.bench_typed export --out typed.jsonl
python -m ollaya_convert.bench_typed run --model d3:mini \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/d3-mini
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model d3:mini \
--url http://127.0.0.1:11435 --data public --out public/d3-mini
# speed and memory
python -m ollaya_convert.bench_latency --models d3:mini --memory \
--machine my-machine --device cuda --out latency.json