Endy-Qwen3.6-CyberSec-35B-A3B

Endy-Qwen3.6-CyberSec-35B-A3B (abliterated, vision) — fp16 weights

QLoRA fine-tune of huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated, specialized for coding, IT and cybersecurity, kept uncensored. Full-precision (fp16) merged weights.

• Architecture: qwen3_5_moe — MoE (~35B total, ~3B active) + linear-attention (DeltaNet) + native MTP + vision (multimodal). Merged directly so the MTP head + vision tower are intact.
• GGUF quantizations (Q2_K…Q8_0 + mmproj for llama.cpp): endystrike/Endy-Qwen3.6-CyberSec-35B-A3B-GGUF
• Use these fp16 weights to run with transformers / vLLM (incl. image input) or to produce your own quantizations.

Training

QLoRA (Unsloth, 4-bit NF4, r32 α64 on q/k/v/o_proj), 2 epochs, train_on_responses_only, on 90,470 coding+cybersecurity chat examples (+914 val). Checkpoint step 2250 selected by validation loss. LoRA merged directly into the fp16 base (preserving MTP + vision), keeping the model multimodal.

Datasets (examples used, licenses)

License

AGPL-3.0 (one training dataset, lordx64/agentic-distill-fable-5-sft, is AGPL-3.0 copyleft → the derived model inherits it).

Disclaimers

• Parts of the base lineage were distilled from proprietary models (Claude Opus 4.7 / Fable 5, DeepSeek V4) by third-party authors; their usage policies may restrict training competing models on their outputs. Disclosed, not waived.
• Uncensored/abliterated — outputs are unfiltered. Intended for authorized security research, pentesting, secure-coding and education. You are responsible for lawful use.
• Not affiliated with or endorsed by Qwen, Anthropic, DeepSeek, or the dataset authors.