d3:edge
Decision 3.0 (d3) · vLLM Semantic Router · 0.6B · ONNX Runtime · Apache-2.0
- Score
- 42.9
- 95 % interval 40.7–45.0
- Rank
- 37
- of 66 scored models
- Public benchmark
- 63.2
- macro accuracy, %
- typed-decisions
- 42.3
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
15.4
Classification
MASSIVE en
78.2
MASSIVE de
60.7
Civil Comments
75.3
Aegis 2
10.4
Understanding
MultiNLI
68.4
VitaminC
60.7
PAWS
65.6
Knowledge
BoolQ
45.3
SQuAD 2
27.8
PubMedQA
22.6
Judging
HelpSteer 2
15.7
SummEval rel.
29.7
SummEval cons.
25.4
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 63.2(61.7–64.7) | 68.7 | 72.4 | 38.9 | 0.092 | 0.459 | 0 / 3,880 |
| typed-decisions | 42.3(39.8–44.6) | 36.7 | 69.0 | 26.4 | 0.345 | 0.859 | 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 | 27.7 ms | 37.3 ms | 6.88 s | 2.6 GB | 3.0 GB | F32 |
| Threadripper 3960X CPU | 1.25 s | 1.26 s | 6.38 s | – | 2.3 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 | 2.0e-5 | pass | results/runs/2026-10-23-parity-rtx5090-cpu.json |
| RTX 5090 CUDA | 696/696 | 6.2e-5 | pass | results/runs/2026-10-23-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:edge
# 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:edge \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/d3-edge
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model d3:edge \
--url http://127.0.0.1:11435 --data public --out public/d3-edge
# speed and memory
python -m ollaya_convert.bench_latency --models d3:edge --memory \
--machine my-machine --device cuda --out latency.json