locateanything-3b-bf16-mlx

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]"

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

• mlx-cv source
• App Automaton
• Upstream model