Blackfrost-AI/PINQWEN-3.6-35B-CLEAN-BF16

PINQWEN-3.6-35B-CLEAN-BF16: A Specialized MoE Assistant

PINQWEN-3.6-35B-CLEAN-BF16 is a supervised fine-tune (SFT) by Blackfrost-AI, based on Alibaba's Qwen/Qwen3.6-35B-A3B.

Key Characteristics & Capabilities

• Architecture: Utilizes a qwen3_5_moe Mixture-of-Experts architecture with approximately 36.97 billion total parameters and 3 billion active parameters per token. It features a hybrid Gated-DeltaNet and gated attention mechanism.
• Fine-tuning: SFT was performed on "The Void (v4)", Blackfrost's proprietary corpus of 5,032 high-quality multi-turn examples, focusing on distilled reasoning, chain-of-thought, and ReAct-style agentic tool-use trajectories.
• Vision-Language: Inherits a full vision tower from the base Qwen3.6-35B model, supporting images and video alongside text. However, Blackfrost-AI's fine-tuning was text-only, meaning vision capabilities track the untuned base model.
• Safety-Aligned: This "CLEAN" variant preserves the base model's safety alignment.
• Training Method: Fine-tuned using bf16 LoRA with Unsloth, applying adapters to attention projections and MoE expert projections, resulting in 1.86 billion trainable parameters.

Intended Use Cases

• General-purpose text assistant: Excels in conversational AI.
• Reasoning tasks: Designed for complex logical deduction.
• Agentic / Tool-use (ReAct-style) tasks: Capable of planning and executing actions using external tools.