Endy-Qwen3.6-CyberSec-35B-A3B — GGUF (abliterated, vision)
GGUF quantizations of a QLoRA fine-tune of huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated, specialized for coding, IT and cybersecurity, kept uncensored, for fast local inference on a single 32 GB GPU (e.g. Tesla V100).
- Architecture:
qwen3_5_moe— MoE (~35B total, ~3B active) + linear-attention (DeltaNet) + native MTP + vision. MTP head is preserved in these GGUFs (block_count 41). - Vision works via the paired
mmprojfile (see below). ~57 tok/s generation on a V100.
Quantizations
| File | Size | Notes |
|---|---|---|
Endy-Qwen3.6-CyberSec-35B-A3B-Q8_0.gguf | 37.8 GB | max fidelity, needs >32 GB VRAM |
Endy-Qwen3.6-CyberSec-35B-A3B-Q6_K.gguf | 29.2 GB | near-lossless |
Endy-Qwen3.6-CyberSec-35B-A3B-Q5_K_M.gguf | 25.3 GB | recommended for 32 GB GPUs |
Endy-Qwen3.6-CyberSec-35B-A3B-Q4_K_M.gguf | 21.7 GB | max context headroom on 24-32 GB |
Endy-Qwen3.6-CyberSec-35B-A3B-Q3_K_M.gguf | 17.2 GB | 20 GB cards, degraded |
Endy-Qwen3.6-CyberSec-35B-A3B-Q2_K.gguf | 13.2 GB | 12-16 GB cards, low precision — weak on coding |
Endy-Qwen3.6-CyberSec-mmproj-f16.gguf | 0.9 GB | vision projector — pair with any quant for image input |
Inference (llama.cpp)
llama-server -m Endy-Qwen3.6-CyberSec-35B-A3B-Q5_K_M.gguf \
--mmproj Endy-Qwen3.6-CyberSec-mmproj-f16.gguf \
-c 262144 -ngl 999 -fa on -ctk q8_0 -ctv q8_0 --jinja
--mmprojenables screenshot/image input; the server then advertises vision so OpenAI-compatible clients send images.- Anti-repetition (recommended) — this model class can loop on long agentic tasks; add server-side (clients can't override these):
--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --repeat-penalty 1.1 --repeat-last-n 512 - Speculative decoding / prefix-KV-reuse are not supported (recurrent linear-attention state can't roll back).
Training (summary)
QLoRA (Unsloth, 4-bit NF4, r32 α64 on q/k/v/o_proj), 2 epochs, train_on_responses_only, on ~90.5k coding+cybersecurity chat examples. Checkpoint step 2250 selected by validation loss. LoRA merged directly into the fp16 base (preserving the MTP head + vision tower), then converted and quantized with llama.cpp.
Datasets (examples used, licenses)
Merged from 12 candidate distill datasets → deduped to 8 unique on-topic sources → 90,470 chat-format examples (+ 914 held-out for validation). Per-source example counts:
| Dataset | Examples | Domain | License | Teacher |
|---|---|---|---|---|
| AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1 | 39,286 | cybersecurity | Apache-2.0 | — |
| Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset | 22,677 | cybersecurity | Apache-2.0 | — |
| WithinUsAI/fable_5_distillation_merged_cleaned_25k | 12,464 | coding | Apache-2.0 | Claude Fable 5 |
| Jackrong/DeepSeek-V4-Distill-8000x | 6,292 | coding | MIT | DeepSeek-V4 |
| lordx64/agentic-distill-fable-5-sft | 4,593 | agentic coding | AGPL-3.0 | Claude Fable 5 |
| CyberNative/Code_Vulnerability_Security_DPO | 4,111 | secure coding | Apache-2.0 | DeepSeek-Coder-33B |
| beyoru/Deepseek-v4-pro-max-distill-1500x | 946 | coding | Apache-2.0 | DeepSeek-V4 |
| WithinUsAI/claude_mythos_distilled (stripped) | 101 | reasoning | Apache-2.0 | declared synthetic |
| Total | 90,470 |
Note: claude_mythos was 25k rows but inflated (~135 unique prompts repeated ~185×) → stripped to 101 representative rows. Malware-source-generation data was deliberately excluded; vulnerability-analysis / pentest / secure-coding kept.
License
AGPL-3.0. One training dataset (lordx64/agentic-distill-fable-5-sft) is AGPL-3.0 (strong copyleft); the derived model inherits AGPL-3.0.
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.