CaptchaKraken_v1.1

A multimodal LoRA adapter for Qwen/Qwen3.5-9B that solves image captchas. It is the model behind CaptchaKraken — a self-hosted captcha solver for browser automation (npm captcha-kraken-js, PyPI captchakraken).

What's new in v1.1

v1 covered grid captchas only ("select all squares with…"). v1.1 extends coverage to the non-grid puzzle families as well — drag-to-slot, click-the-object, line-tracing, tile-fitting, and the other hCaptcha interaction types — while keeping the grid performance. It is trained with per-object content supervision (each click/drag target is labelled by what it is, not just where), which is what recovered the drag/click puzzles.

Details

• Base model: Qwen/Qwen3.5-9B (natively multimodal)
• Adapter: PEFT LoRA, r=32, lora_alpha=64, applied to both the language model and the vision tower (so serving requires vLLM's --enable-tower-connector-lora).
• Task: image captcha solving. Grid tile selection (reCAPTCHA 3×3 / 4×4, hCaptcha 3×3 image grids) plus non-grid puzzles (drag / click / line / fit). For grids it returns the cell numbers to click; for interaction puzzles it returns the action (click points or drag vectors).

Serving (vLLM)

vllm serve Qwen/Qwen3.5-9B \
  --reasoning-parser qwen3 \
  --enable-lora --enable-tower-connector-lora \
  --max-lora-rank 64 --max-model-len 8192 \
  --trust-remote-code --port 8000 \
  --lora-modules captcha=<user>/CaptchaKraken_v1.1

--enable-tower-connector-lora is required — without it the vision half of the adapter is dropped and grid accuracy collapses. On limited-VRAM GPUs, serve a quantized base (AWQ/FP8) instead of the bf16 base above.

The CaptchaKraken project wires this up hands-off: setup.sh downloads the weights and the server auto-starts on the first solve.

License

GPL-3.0-or-later.

Framework versions

• PEFT 0.18.1