Qwen3.6-35B-A3B-NVFP4

Qwen3.6-35B-A3B-NVFP4 is NVIDIA's official NVFP4-quantized checkpoint of Alibaba's Qwen3.6-35B-A3B, a sparse Mixture-of-Experts model with 35B total parameters and 3B activated per token. Quantized via NVIDIA Model Optimizer (modelopt v0.44.0), it compresses the BF16 checkpoint from ~67 GB to ~19 GB on disk (~3.06× reduction) while targeting near-lossless accuracy on NVIDIA Hopper and Blackwell GPUs. The architecture combines 256 experts (8 routed + 1 shared), hybrid attention (Gated DeltaNet + Gated Attention), native 262K context, Multi-Token Prediction (MTP), and multimodal support for text, image, and video inputs.

Official NVIDIA evaluation shows NVFP4 tracks BF16 closely across reasoning, agentic, and multimodal benchmarks: MMLU Pro 85.0 vs 85.6, GPQA Diamond 84.8 vs 84.9, SciCode 40.6 vs 40.8, AIME 2025 88.8 vs 89.2, and MMMU Pro 74.5 vs 74.1. Community NVFP4 preservation studies on the same base model report GSM8K accuracy of ~96.3% with no statistically significant degradation versus FP8 baselines. The model is optimized for production deployment via vLLM with reasoning-parser qwen3, tool-call support, FP8 KV cache, and speculative MTP decoding.

Released on Hugging Face on 2026-05-28 under Apache 2.0, it is widely regarded as one of the best options for running Qwen3.6 MoE models on single Blackwell or Hopper GPUs, offering strong agentic coding and reasoning capabilities at the compute cost of a ~3B active-parameter model. Typical VRAM usage is ~22 GB for weights alone, rising to ~24 GB at 16K context with KV cache overhead.