nimble:9b-v3
Nimble · Bespoke Labs · 9B · ONNX Runtime · CC-BY-NC-4.0 (adapter) and Apache-2.0 (base model)
- Score
- 62.4
- 95 % interval 60.5–64.4
- Rank
- 3
- of 41 scored models
- Public benchmark
- 75.4
- macro accuracy, %
- typed-decisions
- 64.8
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
48.4
Classification
MASSIVE en
85.5
MASSIVE de
83.1
Civil Comments
78.0
Aegis 2
48.0
Understanding
MultiNLI
83.4
VitaminC
69.2
PAWS
63.2
Knowledge
BoolQ
72.0
SQuAD 2
62.5
PubMedQA
62.2
Judging
HelpSteer 2
26.2
SummEval rel.
21.9
SummEval cons.
70.5
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 75.4(74.1–76.7) | 82.7 | 82.4 | 51.6 | 0.092 | 0.334 | 0 / 3,880 |
| typed-decisions | 64.8(62.5–67.0) | 64.5 | 74.7 | 57.5 | 0.195 | 0.515 | 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 | 942 ms | 947 ms | 11.4 s | 20.7 GB | 3.1 GB | F32 |
| Threadripper 3960X CPU | 50.9 s | 51.1 s | 19.1 s | – | 18.8 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 4090 CUDA | 498/498 | 1.4e-4 | pass | results/runs/2026-10-19-parity-rtx4090-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 nimble:9b-v3
# 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 nimble:9b-v3 \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/nimble-9b-v3
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model nimble:9b-v3 \
--url http://127.0.0.1:11435 --data public --out public/nimble-9b-v3
# speed and memory
python -m ollaya_convert.bench_latency --models nimble:9b-v3 --memory \
--machine my-machine --device cuda --out latency.json