This build uses the existing MLX language weights from KCh3dRi4n/Qwen3.6-35B-A3B-NVIDIA-NVFP4 and restores the missing vision_tower.* weights plus the image/video processor configs from the NVIDIA checkpoint. It is intended for mlx_vlm, not text-only mlx_lm.

Quantization note: the language/MoE weights remain MLX NVFP4. The restored vision tower is BF16, matching the upstream NVIDIA checkpoint where the vision weights are not NVFP4-quantized.

Context: 512K YaRN context.

bash

pip install mlx-vlm
python -m mlx_vlm.generate --model KCh3dRi4n/Qwen3.6-35B-A3B-NVIDIA-NVFP4-512K --image /path/to/image.jpg --prompt "Describe this image."