d3:nano
Decision 3.0 (d3) · vLLM Semantic Router · 2B · ONNX Runtime · Apache-2.0
- Score
- 62.2
- 95 % interval 60.3–64.3
- Rank
- 5
- of 66 scored models
- Public benchmark
- 75.6
- macro accuracy, %
- typed-decisions
- 62.5
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
45.0
Classification
MASSIVE en
83.4
MASSIVE de
81.2
Civil Comments
86.0
Aegis 2
40.8
Understanding
MultiNLI
82.9
VitaminC
70.7
PAWS
70.4
Knowledge
BoolQ
75.3
SQuAD 2
47.8
PubMedQA
51.4
Judging
HelpSteer 2
29.2
SummEval rel.
34.9
SummEval cons.
72.2
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 75.6(74.2–76.9) | 80.7 | 82.0 | 56.4 | 0.062 | 0.325 | 0 / 3,880 |
| typed-decisions | 62.5(60.0–64.8) | 63.3 | 75.2 | 52.3 | 0.117 | 0.514 | 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 | 88.9 ms | 92.2 ms | 30.2 s | 8.2 GB | 11.4 GB | F32 |
| Threadripper 3960X CPU | 3.93 s | 3.96 s | 33.5 s | – | 11.1 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 | 1.5e-5 | pass | results/runs/2026-10-24-parity-rtx5090-cpu.json |
| RTX 5090 CUDA | 696/696 | 1.9e-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:nano
# 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:nano \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/d3-nano
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model d3:nano \
--url http://127.0.0.1:11435 --data public --out public/d3-nano
# speed and memory
python -m ollaya_convert.bench_latency --models d3:nano --memory \
--machine my-machine --device cuda --out latency.json