Dark-Pane v2

Status: EXPERIMENT. Public Dark-Pane family checkpoint.

Abliterated Qwen3.8-27B for Apple Silicon (MLX). Native vision-language, 262k context, 2 / 4 / 6 / 8-bit.

This is a rehost of orcarouter/Qwen3.8-27B-Uncensored-MLX into the Dark-Pane family. Weights are OrcaRouter's abliterated MLX quants of Qwen's Apache-2.0 27B VL model — not the Gemma 4 12B SOM+ASPA surgery from Dark-Pane.

Repo root = 4-bit. Other precisions live in 2-bit/, 4-bit/, 6-bit/, 8-bit/.

---

⚠️ Research use

Safety alignment has been substantially removed (abliteration). It will comply with requests the stock Qwen3.8-27B would refuse. Intended for alignment research, red-teaming, and local evaluation. You are responsible for what you generate. Add your own guardrails before any deployment.

Apache 2.0 from the Qwen base. Attribution: Qwen / Alibaba (base), OrcaRouter (abliteration + MLX quants).

---

Quants

Vision tower, norms, and conv layers stay BF16. Only LM linear weights are quantized.

---

Usage (mlx-vlm)

pip install -U mlx-vlm    # mlx-vlm >= 0.6.13, mlx >= 0.32

hf download mars2titan/darkpane-v2 --include "4-bit/*" --local-dir ./darkpane-v2

python -m mlx_vlm generate \
  --model ./darkpane-v2/4-bit \
  --prompt "Explain quantum entanglement in one sentence." --max-tokens 256

python -m mlx_vlm generate \
  --model ./darkpane-v2/4-bit \
  --image path/to/image.png \
  --prompt "Describe this image." --max-tokens 256

python -m mlx_vlm server --model ./darkpane-v2/4-bit --port 8080

LM Studio

Search mars2titan/darkpane-v2. Root is 4-bit. Turn off KV cache quantization (MLX vision on this arch fails if it is on). 4-bit fits a 32 GB Mac; 8-bit wants 64 GB.

---

Model