d3:lite
Decision 3.0 (d3) · vLLM Semantic Router · 0.8B · ONNX Runtime · Apache-2.0
- Score
- 49.9
- 95 % interval 47.9–51.8
- Rank
- 31
- of 66 scored models
- Public benchmark
- 67.0
- macro accuracy, %
- typed-decisions
- 52.6
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
30.6
Classification
MASSIVE en
83.7
MASSIVE de
80.6
Civil Comments
78.7
Aegis 2
18.4
Understanding
MultiNLI
81.9
VitaminC
66.9
PAWS
68.8
Knowledge
BoolQ
62.7
SQuAD 2
45.8
PubMedQA
40.6
Judging
HelpSteer 2
15.2
SummEval rel.
38.0
SummEval cons.
-13.7
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 67.0(65.6–68.3) | 78.5 | 77.4 | 30.5 | 0.077 | 0.381 | 0 / 3,880 |
| typed-decisions | 52.6(50.2–55.0) | 50.7 | 66.3 | 43.9 | 0.134 | 0.610 | 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 | 65.2 ms | 68.4 ms | 9.43 s | 4.4 GB | 5.4 GB | F32 |
| Threadripper 3960X CPU | 2.95 s | 2.98 s | 12.7 s | – | 4.6 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 | 4.7e-5 | pass | results/runs/2026-10-24-parity-rtx5090-cpu.json |
| RTX 5090 CUDA | 696/696 | 5.4e-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:lite
# 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:lite \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/d3-lite
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model d3:lite \
--url http://127.0.0.1:11435 --data public --out public/d3-lite
# speed and memory
python -m ollaya_convert.bench_latency --models d3:lite --memory \
--machine my-machine --device cuda --out latency.json