Qwen3-4B ReAct Browser Agent
Fine-tuned Qwen3-4B for on-device browser agent tasks (TinyBrowser/Wiegand).
Training
- SFT: MLX LoRA, rank=16, all layers + lm_head, 571 examples, loss 2.26→0.015
- Base: Qwen/Qwen3-4B-Instruct-2507
- Task: ReAct loop JSON tool calls (click, type_and_wait, scroll, extract_and_done, etc.)
Files
qwen3-4b-react-q4_0.gguf — Q4_0 quantized GGUF for llama.cpp (Adreno OpenCL)mlx-lora/ — MLX LoRA adapter weightsdata/qwen3_react.jsonl — 571 training examples
Usage
Drop the GGUF on device at /sdcard/models/ and select in Settings → Model Manager.
