Unsloth Releases NVFP4 Quantization for Qwen3.6

Unsloth has unveiled new NVFP4 Quantization models for Qwen3.6. These models deliver up to 2.5x faster performance compared to before when run on NVIDIA's Blackwell architecture (e.g., RTX 5090, DGX Spark, etc.).

Key features are as follows:

• Inference Performance: Based on the Qwen3.6-35B-A3B model, it recorded a speed of 17,561 tok/s on a B200 GPU.
• Memory Efficiency: The Qwen3.6-27B NVFP4 model can run in a 24GB VRAM environment.
• Accuracy Improvements: The quantization process improved Accuracy, Tool calling, Agent use, and Looping capabilities.

The quantized models have been distributed via Hugging Face, and for optimal performance, using a Blackwell GPU along with following the guide provided by Unsloth is required.