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

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:agpl-3.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Endy-Qwen3.6-CyberSec-35B-A3B by endystrike is a QLoRA fine-tuned Qwen3.6-35B-A3B model, featuring a MoE architecture (~35B total, ~3B active) with linear-attention and native MTP, plus integrated vision capabilities. Specialized for coding, IT, and cybersecurity, this uncensored model is designed for tasks requiring deep technical understanding in these domains. It was trained on over 90,000 coding and cybersecurity chat examples, making it highly proficient for security research, pentesting, and secure-coding applications.

Loading preview...

Model Overview

Endy-Qwen3.6-CyberSec-35B-A3B is a QLoRA fine-tune by endystrike, built upon the huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated base model. This model utilizes a qwen3_5_moe architecture, featuring a Mixture-of-Experts (MoE) design with approximately 35 billion total parameters and 3 billion active parameters. It incorporates linear-attention (DeltaNet), native MTP, and crucially, vision capabilities, making it a multimodal model.

Key Specializations

This model is specifically specialized for coding, IT, and cybersecurity tasks. It has been kept uncensored to facilitate its intended use cases. The fine-tuning process involved 2 epochs of QLoRA training (4-bit NF4, r32 α64 on q/k/v/o_proj) on a substantial dataset of 90,470 coding and cybersecurity chat examples.

Training Data & Licensing

The training datasets include a diverse range of cybersecurity and coding-focused examples, such as AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1, Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset, and various distillation datasets from models like Claude Fable 5 and DeepSeek-V4. Notably, one dataset, lordx64/agentic-distill-fable-5-sft, carries an AGPL-3.0 license, which the derived model inherits.

Intended Use Cases

  • Authorized security research
  • Penetration testing
  • Secure-coding development
  • Cybersecurity education

Users should be aware that the model is uncensored and outputs are unfiltered, requiring responsible and lawful use. The model's base lineage includes distillation from proprietary models, which may have usage policy implications.

Finetunes
2 models