NVIDIA Releases Qwen3.6 NVFP4

NVIDIA has released a quantized version of Alibaba's Qwen3.6-35B-A3B model using the NVFP4 data type.

This model applies Post Training Quantization using NVIDIA Model Optimizer, and is optimized for inference via vLLM.

Key Features:

• Memory Reduction: By reducing bits per parameter from 16-bit to 4-bit, disk and GPU memory requirements were reduced by approximately 3.06x.
• Quantization Scope: Quantization was applied to the weights and activation functions of linear operators in the transformer blocks within the MoE (Mixture of Experts) architecture.
• Performance Retention: The model maintained high accuracy with very minimal performance degradation compared to BF16 on major benchmarks including MMLU Pro (85.0) and GPQA Diamond (94.7).