Qwen3.6-35B-A3B-4bit-Native This repository provides the 4-bit (NF4) quantized weights for the Qwen3.6-35B-A3B Mixture-of-Experts model. These weights were generated using the bitsandbytes library with double quantization enabled to ensure maximum precision at a reduced memory footprint.
Model Details
Base Model: Qwen3.6-35B-A3B Quantization: 4-bit NormalFloat (NF4) Framework: Hugging Face Transformers Total Parameters: ~35B Expert Architecture: 256 Experts per Layer
Key Features
Native Compatibility: Designed to work seamlessly with the transformers library without additional conversion layers. Memory Efficiency: Optimized to fit within ~20GB of memory (VRAM/RAM combined), making it accessible for mid-range hardware environments. Precision: Uses Double Quantization to minimize perplexity degradation compared to the original BF16 weights.