File	Size	Notes
`onnx/model_quantized.onnx`	619 MB	int8 dynamic-quantized decoder (self-contained, no external data)
`config.json` / `tokenizer*.json`	—	Qwen3 config + tokenizer (chat template embedded)

Usage (browser / Node)

import { pipeline } from '@huggingface/transformers';

const generator = await pipeline('text-generation', 'Luigi/dict-xs-onnx', {
  dtype: 'q8',      // CPU/WASM
  // device: 'webgpu', dtype: 'q4f16'  // optional GPU path
});

const out = await generator([
  { role: 'system', content: 'You are a bilingual dictionary...' },
  { role: 'user', content: 'Expression: "maison"\nContext: ...' },
], { max_new_tokens: 50, temperature: 0.3, top_p: 0.9 });

Conversion notes

Exported from ikhou/dict-xs (bfloat16 safetensors, step 6568 = the repo's main weights) into the fused decoder graph (GroupQueryAttention + RotaryEmbedding contrib ops, opset 18, position_ids input) used by onnx-community/Qwen3-0.6B-ONNX, with the fine-tuned weights transplanted.
Numerically verified against the PyTorch checkpoint (max logit diff ≈ 5e-5 fp32; greedy decode matches the HF reference up to int8 noise).
Generated with the repo's own export script: scripts/export_onnx.py (see ikhou/dict-xs-demo space for the browser demo).