d3:flash
Decision 3.0 (d3) · vLLM Semantic Router · 9B · ONNX Runtime · Apache-2.0
- Score
- 67.8
- 95 % interval 66.0–69.5
- Rank
- 2
- of 66 scored models
- Public benchmark
- 79.0
- macro accuracy, %
- typed-decisions
- 71.2
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
57.8
Classification
MASSIVE en
88.8
MASSIVE de
88.2
Civil Comments
85.3
Aegis 2
59.2
Understanding
MultiNLI
90.0
VitaminC
74.7
PAWS
76.8
Knowledge
BoolQ
79.3
SQuAD 2
44.5
PubMedQA
59.2
Judging
HelpSteer 2
32.2
SummEval rel.
32.3
SummEval cons.
80.9
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 79.0(77.7–80.2) | 85.5 | 84.5 | 58.8 | 0.030 | 0.274 | 0 / 3,880 |
| typed-decisions | 71.2(68.7–73.4) | 73.5 | 77.2 | 65.0 | 0.032 | 0.383 | 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 | 317 ms | 320 ms | 39.1 s | 18.0 GB | 2.9 GB | F32 |
| Threadripper 3960X CPU | 14.8 s | 14.9 s | 19.7 s | – | 16.0 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 | 1.1e-4 | pass | results/runs/2026-10-24-parity-rtx5090-cpu.json |
| RTX 5090 CUDA | 696/696 | 2.9e-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:flash
# 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:flash \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/d3-flash
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model d3:flash \
--url http://127.0.0.1:11435 --data public --out public/d3-flash
# speed and memory
python -m ollaya_convert.bench_latency --models d3:flash --memory \
--machine my-machine --device cuda --out latency.json