kev:9b
Kev · Jared Palmer · 7.9B · ONNX Runtime · Apache-2.0
- Score
- 62.6
- 95 % interval 60.6–64.5
- Rank
- 5
- of 66 scored models
- Public benchmark
- 75.3
- macro accuracy, %
- typed-decisions
- 72.2
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 75.3(73.9–76.6) | 81.6 | 80.7 | 55.6 | 0.058 | 0.356 | 0 / 3,880 |
| typed-decisions | 72.2(69.7–74.5) | 69.3 | 81.2 | 67.5 | 0.095 | 0.410 | 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 | 248 ms | 250 ms | 39.1 s | 17.9 GB | 2.9 GB | F32 |
| Threadripper 3960X CPU | 12.5 s | 12.5 s | 39.2 s | – | 15.9 GB | F32 |
kev:9b-fp16 RTX 5090 | 125 ms | 126 ms | 15.2 s | 17.6 GB | 3.0 GB | F16 |
kev:9b-fp16 is the same model computing in fp16 on the GPU, not ranked separately. Its decisions agree with kev:9b's on at least 99.5 % of the parity questions; typed-decisions 72.3 against 72.2, measured with Ollaya 0.14.0.
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 |
|---|---|---|---|---|
| i9-13900K; ONNX Runtime 1.28 | 100 % | 3.6e-5 | pass | families/kev.md:242 |
| RTX 4090 CUDA | 480/480 | 7.5e-5 | 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 kev:9b
# 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 kev:9b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/kev-9b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model kev:9b \
--url http://127.0.0.1:11435 --data public --out public/kev-9b
# speed and memory
python -m ollaya_convert.bench_latency --models kev:9b --memory \
--machine my-machine --device cuda --out latency.json