CaptchaKrakenV1_Lora

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).

This adapter supersedes the earlier CaptchaKraken_v1 and CaptchaKraken_v1.1 repositories, which have been withdrawn. It is the same trained adapter as v1.1, republished under the GNU Lesser General Public License, version 3 only, and under a stable name.

Coverage

The first release covered grid captchas only ("select all squares with…"). This adapter also handles the non-grid puzzle families — drag-to-slot, click-the-object, line-tracing, tile-fitting, and the other 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), revision c202236235762e1c871ad0ccb60c8ee5ba337b9a — the revision this adapter was trained against.
• Adapter: PEFT LoRA, r=32, lora_alpha=64, applied to both the language model and the vision tower, excluding the vision merger. Serving the adapter therefore requires vLLM's --enable-tower-connector-lora.
• Task: image captcha solving. Grid tile selection (reCAPTCHA 3×3 / 4×4 and other 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 32 --max-model-len 8192 \
  --trust-remote-code --port 8000 \
  --lora-modules captcha=CaptchaKraken/CaptchaKrakenV1_Lora

--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, or use one of the pre-merged single-file models below.

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

Pre-merged models

If you would rather not wire up an adapter, this adapter merged into the base is published as two standalone models — no adapter flags, one download:

Model	Precision	Weights	Min VRAM
CaptchaKraken/Sunlight-AWQ-4bit	4-bit (AWQ)	~6 GB	11 GB
CaptchaKraken/Twilight-FP8	8-bit (FP8)	~14 GB	22 GB

License

LGPL-3.0-only — see LICENSE. Use, modify and redistribute these weights, including in commercial and hosted services. If you distribute a modified version, keep the licence and copyright notices and make your changes available under LGPLv3.

Framework versions

• PEFT 0.18.1