Exported the query safety classification heads to ONNX.
The ONNX graph accepts input_ids and attention_mask.
The graph returns risk_logits with three classes and category_logits with nine classes.
Replaced the upstream export-time causal mask implementation with an equivalent explicit additive causal mask because the original vmap implementation could not be traced by the ONNX exporter.
Applied signed INT8 dynamic quantization, including the token embedding table.
Runtime input is limited to 512 tokens while preserving the complete Qwen user-message prefix and suffix.

The exported logits were checked against the original PyTorch model before quantization. Across the local qualification corpus, signed INT8 produced no Safe versus non-Safe boundary changes. Some Unsafe versus Controversial tier changes were observed.

Files

model.onnx: ONNX graph.
model.onnx.data: external signed INT8 model weights.
tokenizer.json: tokenizer from the pinned upstream revision.
LICENSE: upstream Apache License 2.0 text.

Checksums

text

83b01b5df2d570089918da04592d9e32022d357e8d18044d9f46d8cdbb1aa47a  model.onnx
d105fd59923cb917194b1efcb7460b771aefd5c2eb5015f4b9948b11d07e82de  model.onnx.data
aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4  tokenizer.json

Local qualification

Tests were run on Apple Silicon with ONNX Runtime using 100 multilingual and mixed-length inputs between 20 and 300 characters:

text

Average: 42.86 ms
P50:     44.39 ms
P95:     64.13 ms
P99:     69.33 ms
Maximum: 70.28 ms

The production runtime uses the final token from the query risk and query category outputs. These measurements are qualification results, not guarantees for other operating systems or hardware.

License

The upstream model and this derivative are distributed under the Apache License 2.0. See LICENSE. Qwen and Alibaba do not endorse this derivative.

NeuraStudios/clipit-qwen3guard-stream-0.6b-onnx-int8

Clip It Qwen3Guard Stream 0.6B ONNX INT8

Source

Modifications

Files

Checksums

Local qualification

License

NeuraStudios/
clipit-qwen3guard-stream-0.6b-onnx-int8