qwen3guard:0.6bbuilt-in questions
Qwen3Guard · Qwen · 0.6B · ONNX Runtime · Apache-2.0
This model answers only its own built-in questions, so the accuracy tasks do not apply. Its speed and memory are below.
Speed and memory
| Device | p50 | p90 | Load | VRAM | RAM peak | Precision |
|---|---|---|---|---|---|---|
| RTX 5090 CUDA | 30.2 ms | 32.3 ms | 10.3 s | 3.6 GB | 5.4 GB | F32 |
| Threadripper 3960X CPU | 1.49 s | 1.51 s | 9.62 s | – | 5.0 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 |
|---|---|---|---|---|
| 24 cores; ONNX Runtime 1.28 | 100 % (296 / 296) | 5.7e-5 | pass | families/qwen3guard.md:157-163 |
| RTX 4090, CUDA 13; ONNX Runtime 1.28 | 100 % (296 / 296) | 7.8e-5 | pass | families/qwen3guard.md:157-163 |
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 qwen3guard:0.6b
# 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 qwen3guard:0.6b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/qwen3guard-0.6b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model qwen3guard:0.6b \
--url http://127.0.0.1:11435 --data public --out public/qwen3guard-0.6b
# speed and memory
python -m ollaya_convert.bench_latency --models qwen3guard:0.6b --memory \
--machine my-machine --device cuda --out latency.json