--- license: apache-2.0 base_model: - Qwen/Qwen3-0.6B-Base - YixuanEvenXu/Qwen3-0.6B-Base-HIP-adapter pipeline_tag: text-generation language: - en tags: - onnx - onnxruntime-web - webgpu - paraphrasing - humanizer --- # Qwen3-0.6B HIP humanizer (web, 4-bit ONNX) A 4-bit ONNX build of the **HIP** paraphraser from *Base Models Look Human To AI Detectors* (Xu, Zhong, Raghunathan, Fang, Kolter, Carnegie Mellon University, 2026), made to run in a web browser with onnxruntime-web on WebGPU or the processor (WASM). It rewrites AI-sounding text into plainer, more natural prose. Test it for free here : https://utilitydone.com/ai-humanizer ## What was changed - The released LoRA adapter `YixuanEvenXu/Qwen3-0.6B-Base-HIP-adapter` was merged into `Qwen/Qwen3-0.6B-Base`. - Converted with onnxruntime-genai 0.20 (`-p int4 -e cpu`, `block_size=32`, MatMul and Gather quantized, embeddings shared with the LM head), max positions 8192. - The 4-bit `GatherBlockQuantized` embedding lookup was replaced by standard ONNX ops (Gather, BitShift, Sub, Cast, Mul) reading the same packed weights, because onnxruntime-web's WASM backend has no kernel for it. Outputs are identical (max logit difference 0.0). - Saved as a single `model.onnx` (377 MB). ## Use Prompt format (no chat template): ``` {text} ``` Generate until `` or end of text. Settings used: temperature 0.7, top_p 0.9, repetition penalty 1.05, one pass. Inputs: `input_ids`, `attention_mask`, `past_key_values.{0-27}.{key,value}` (float32, `[1, 8, past, 128]`); outputs: `logits`, `present.*`. ## Licence and credits Apache-2.0, like the Qwen3 base model and the HIP adapter. Credits: Qwen team (Qwen3-0.6B-Base); Yixuan Even Xu et al. (HIP, [arXiv:2605.19516](https://arxiv.org/abs/2605.19516), [code](https://github.com/YixuanEvenXu/humanization-by-iterative-paraphrasing)).