vanch007/WeMM-Video-35BA3B-mlx-8bit

Apple Silicon MLX 8-bit mixed-precision conversion of tencent/WeMM-Video-35BA3B, built with vanch007/mlx-wemm-video for native MLX and oMLX inference.

Quantization Details

  • Language backbone & 256 MoE experts: 8-bit affine (group_size=64)
  • MoE router & shared_expert_gate: 8-bit affine (group_size=64)
  • Vision tower (27-layer Qwen3-VL ViT): bfloat16 (unquantized)
  • Audio tower (32-layer Whisper + 2-layer BERT Q-Former): float16 (unquantized; verified cos=0.999826 vs FP32 reference)
  • Gated DeltaNet A_log: float32
  • Total disk size: 40.34 GB

Other Variants

Usage

git clone https://github.com/vanch007/mlx-wemm-video.git
cd mlx-wemm-video
python -m mlx_wemm_video.generate \
  --model /path/to/WeMM-Video-35BA3B-mlx-8bit \
  --video examples/media/apollo-11-first-steps.mp4 \
  --prompt "Describe the scene and summarize the speech."
Downloads last month
-
Safetensors
Model size
36B params
Tensor type
F16
路
U32
路
BF16
路
F32
路
MLX
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for vanch007/WeMM-Video-35BA3B-mlx-8bit

Quantized
(7)
this model