CaptchaKraken Sunlight (AWQ 4-bit)

A standalone captcha-solving model: the CaptchaKrakenV1_Lora adapter merged into Qwen/Qwen3.5-9B and quantized to AWQ 4-bit. One download, no adapter to wire up.

Sunlight is the lighter of two hardware tiers. Twilight is the same adapter and base at FP8, ~5 GB larger and ~4 points stronger.

  • Weights: ~9.1 GB Β· Min VRAM: ~11 GB
  • Base revision: c202236235762e1c871ad0ccb60c8ee5ba337b9a
  • Quantized: language-model linears at 4-bit (group 128, asymmetric). The vision tower, the linear-attention projections and the multi-token-prediction head stay bf16 β€” quantizing those costs accuracy out of proportion to the space saved. Calibrated on 256 in-domain captchas (grid and interaction puzzles, none from the held-out set).
  • Solves: image grids, plus drag / click / path / fit puzzles.

Serving (vLLM)

vllm serve CaptchaKraken/Sunlight-AWQ-4bit \
  --max-model-len 8192 --gpu-memory-utilization 0.85 --trust-remote-code --port 8000

No --enable-lora and no adapter flags β€” the adapter is already merged in.

Prompts β€” read this before integrating

This model only performs as measured when it is sent the prompts it was trained on. prompts.json in this repo carries them, along with the puzzle-type β†’ template mapping and a prompt_version. Resolve prompts from that file rather than hardcoding a copy: a mismatched prompt does not error, it silently collapses accuracy.

Two integration requirements:

  • Disable thinking. Send chat_template_kwargs: {"enable_thinking": false}. With thinking on and a qwen3 reasoning parser, the answer is routed into reasoning and content comes back empty.
  • Coordinates are normalized 0–1000, top-left (0,0), bottom-right (1000,1000) β€” not pixels.

Evaluation

Scored with the project's soft-tolerance grader on 156 human-labelled held-out captchas, through an OpenAI-compatible vLLM endpoint. Samples promoted into the labelled-train pool are excluded, so nothing here was trained on. Grid puzzles are exact tile-set match with per-error decay; click/drag receive partial credit by normalized distance β€” so these are graded scores, not "percent solved".

model score
Twilight (FP8) 66.87%
production captcha (FP8 base + LoRA at serve time) 66.30%
Sunlight (AWQ 4-bit) 62.72%

Merging costs nothing: Twilight matches the served base+adapter setup within noise. The 4-bit gap concentrates in coordinate-precision puzzles (drag, path) rather than tile selection, where the two are identical.

License

LGPL-3.0-only β€” 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.

Downloads last month
59
Safetensors
Model size
10B params
Tensor type
I32
Β·
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for CaptchaKraken/Sunlight-AWQ-4bit

Finetuned
Qwen/Qwen3.5-9B
Quantized
(732)
this model

Collection including CaptchaKraken/Sunlight-AWQ-4bit