UI-TARS-1.5-7B-8bit (MLX)

8-bit MLX quantization of ByteDance-Seed/UI-TARS-1.5-7B, a Qwen2.5-VL vision-language GUI-agent model for grounded computer/browser control.

• Quantization: 8-bit affine, group size 64 (~9.11 effective bits/weight)
• Architecture: Qwen2_5_VLForConditionalGeneration (qwen2_5_vl) — ViT vision encoder + Qwen2.5 LM
• Converted with: mlx_vlm.convert (mlx-vlm)

Use with mlx-vlm

pip install mlx-vlm
python -m mlx_vlm generate \
  --model mlx-community/UI-TARS-1.5-7B-8bit \
  --image screenshot.png \
  --prompt "Click the search box." \
  --max-tokens 128

This 8-bit build complements the existing mlx-community/UI-TARS-1.5-7B-4bit and -6bit — higher fidelity at ~8.8 GB.