LocateAnything-3B BF16 for MLX
Final-layout BF16 weights for running NVIDIA LocateAnything-3B with mlx-cv on Apple Silicon. BF16 is reduced precision, not integer quantization.
pip install "mlx-cv[mlx,hub]==0.0.3"
from mlx_cv.models.locateanything import LocateAnythingPipeline
pipeline = LocateAnythingPipeline.from_pretrained("locateanything-3b-bf16")
result = pipeline.predict(image, "find every traffic sign")
Verification and performance
The MLX FP32 port first passed the upstream parameter and selected-tap parity gate. The BF16 package then preserved generated tokens and output geometry on four sequential real-image checks (desktop, street signs, document, and webpage). Local peak-memory observations ranged from roughly 9.8 GB to 52.3 GB depending on image and output complexity; these are machine-specific measurements, not requirements or guarantees.
One desktop multi-category prompt repeatedly emitted a monitor category. This known behavior is recorded as a model/output limitation rather than hidden by post-processing.
Limitations
- Inference only, on MLX-supported Apple Silicon systems.
- Visual grounding output can omit, repeat, or mislabel objects; validate it for consequential uses.
- Latency and memory vary substantially with image resolution, prompt, and requested output density.
- This conversion does not change the upstream acceptable-use or license restrictions.
License
The weights retain the bundled NVIDIA License and are restricted to academic and non-profit research purposes. Commercial use is not permitted except as described by that license. mlx-cv code is MIT licensed separately.
Links
- Downloads last month
- -
Quantized
# Make sure mlx-vlm is installed # pip install --upgrade mlx-vlm from mlx_vlm import load, generate from mlx_vlm.prompt_utils import apply_chat_template from mlx_vlm.utils import load_config # Load the model model, processor = load("appautomaton/locateanything-3b-bf16-mlx") config = load_config("appautomaton/locateanything-3b-bf16-mlx") # Prepare input image = ["http://images.cocodataset.org/val2017/000000039769.jpg"] prompt = "Describe this image." # Apply chat template formatted_prompt = apply_chat_template( processor, config, prompt, num_images=1 ) # Generate output output = generate(model, processor, formatted_prompt, image) print(output)