mlx-community/LocateAnything-3B-4bit

MLX mixed 4/8-bit (mixed_4_8, ~6.7 bits/weight) conversion of nvidia/LocateAnything-3B, a vision-language model for fast, high-quality visual grounding (object detection, referring-expression grounding, pointing, GUI/text localization). Converted with mlx-vlm for Apple Silicon.

Box coordinates stay accurate (within ~1-2 quant levels of bf16); semantic labels may generalize (e.g. object instead of remote). Pure 4-bit was not released because quantizing the tied embed_tokens/lm_head destroys coordinate-token precision.

Requirements

Note: LocateAnything support in mlx-vlm currently lives in a pull request and is not yet in a released mlx-vlm. Until it merges, install from the branch that adds the locateanything model:

pip install "git+https://github.com/beshkenadze/mlx-vlm@feat/locateanything-3b"

Usage

python -m mlx_vlm.generate --model mlx-community/LocateAnything-3B-4bit \
  --image http://images.cocodataset.org/val2017/000000039769.jpg \
  --prompt "Detect all objects in the image." --max-tokens 128 --temperature 0.0

Output is structured coordinate tokens, e.g. <ref>remote</ref><box><64><152><273><244></box> with coordinates quantized to <0>..<1000> (normalized). Decoding modes: autoregressive (slow, default) and Parallel Box Decoding (fast/hybrid, ~2x faster) via generation_mode.

Attribution & license

• Derived from nvidia/LocateAnything-3B — released under the NVIDIA License: non-commercial, research/academic use only (commercial use not permitted except by NVIDIA). Redistribution must retain this license and attribution.
• Vision encoder: MoonViT-SO-400M (MIT). Language model: Qwen2.5-3B-Instruct (Qwen Research License). Part of the Eagle VLM family.

The LICENSE file from the source model is included in this repo.