A standalone captcha-solving model: the CaptchaKrakenV1_Lora adapter merged into Qwen/Qwen3.5-9B and quantized to FP8-dynamic. One download, no adapter to wire up.

Twilight is the stronger of two hardware tiers, and matches the accuracy of serving the base with the adapter attached at runtime. Sunlight is the same adapter and base at AWQ 4-bit, ~5 GB smaller and ~4 points weaker.

Weights: ~14.0 GB · Min VRAM: ~22 GB
Base revision: c202236235762e1c871ad0ccb60c8ee5ba337b9a
Quantized: language-model linears at FP8 (channel-wise weights, per-token dynamic activations). The vision tower, the linear-attention projections and the multi-token-prediction head stay bf16 — the same configuration as the FP8-dynamic base this project serves in production, so numerics match what the hosted API runs. FP8-dynamic needs no calibration data.
Solves: reCAPTCHA and hCaptcha grids, plus hCaptcha drag / click / path / fit puzzles.

Serving (vLLM)

bash

vllm serve CaptchaKraken/Twilight-FP8 \
  --max-model-len 8192 --gpu-memory-utilization 0.85 --trust-remote-code --port 8000

No --enable-lora and no adapter flags — the adapter is already merged in.

Prompts — 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 → 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) — 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 — 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

CaptchaKraken Source-Available License v1.0 — see LICENSE.

Use it, modify it, build commercially on it where captcha solving is an internal component of a product that delivers value beyond the solve — scrapers, anti-detection browsers, automation frameworks, QA and accessibility tooling.

You may not sell the solve: no reselling or offering for a fee a captcha solving service or API whose primary value is solving captchas, no thin wrappers exposing this model's solving capability, and no relaying its outputs through a paid or public captcha-solving API. Those require a separate written commercial agreement.