---
license: apache-2.0
library_name: mlx
tags:
- mlx
- 3d-reconstruction
- depth-estimation
- point-cloud
- vision
- apple-silicon
- metal
pipeline_tag: depth-estimation
base_model: robbyant/lingbot-map
---
# MLX LingBot-MAP Weights

Native Apple Silicon (Metal GPU) Accelerated Weights for **[LingBot-MAP / Geometric Context Transformer (GCT)](https://github.com/Robbyant/lingbot-map)**, converted for [Apple MLX](https://github.com/ml-explore/mlx).

- **Original Project Repository**: [Robbyant/lingbot-map](https://github.com/Robbyant/lingbot-map)
- **Original PyTorch Weights**: [robbyant/lingbot-map](https://huggingface.co/robbyant/lingbot-map)
- **MLX Model Repository**: [uqer1244/mlx_lingbot-map](https://huggingface.co/uqer1244/mlx_lingbot-map)
- **Primary Weight File**: `lingbot-map-fp16.safetensors` (~2.16 GB)
- **License**: Apache License 2.0

---

## 📦 Available Model Weight Variants

We provide pre-converted MLX `safetensors` weights in 5 precision formats:

| Format / Precision | File Name | Size | Target Hardware / Description |
| :--- | :--- | :--- | :--- |
| **FP16 (Recommended)** | `lingbot-map-fp16.safetensors` | **2.16 GB** | **Default recommended precision** for M1/M2/M3/M4 GPUs |
| **FP32** | `lingbot-map-fp32.safetensors` | **4.31 GB** | Full precision reference weights |
| **BF16** | `lingbot-map-bf16.safetensors` | **2.16 GB** | BFloat16 precision for M2 / M3 / M4 Apple Silicon |
| **INT8** | `lingbot-map-int8.safetensors` | **1.24 GB** | 8-bit Group-wise Affine Quantization for low-memory devices |
| **INT4** | `lingbot-map-int4.safetensors` | **0.72 GB** | 4-bit Group-wise Affine Quantization for minimal RAM usage |

---

## 🚀 Quick Usage Guide

### 1. Download via Hugging Face Hub (Python)

```python
from huggingface_hub import hf_hub_download
# Download default FP16 MLX safetensors weight file
weights_path = hf_hub_download(
    repo_id="uqer1244/mlx_lingbot-map",
    filename="lingbot-map-fp16.safetensors",
    local_dir="checkpoints"
)
print(f"Weights downloaded to: {weights_path}")
```

### 2. Load Weights into MLX

```python
import mlx.core as mx
# Load safetensors directly in MLX
weights = mx.load("checkpoints/lingbot-map-fp16.safetensors")
print(f"Loaded {len(weights)} MLX layer tensors!")
```

### 3. Run Streaming 3D Reconstruction Demo

Clone the MLX project repository:

```bash
git clone https://github.com/uqer1244/mlx_lingbot-map.git
cd mlx_lingbot-map
pip install -e .

# Run streaming 3D reconstruction
python create_map.py --image_folder path/to/images --stride 2 --out_map maps/reconstruction_map.npz

# Launch interactive 3D Web Visualizer
python view_map.py --map_file maps/reconstruction_map.npz --port 8080

```

---

## 📜 Acknowledgements & Citation

This model weight repository contains MLX-converted `.safetensors` derived from the original **LingBot-MAP** project created by the Robbyant team:

- Original Code Repository: [https://github.com/Robbyant/lingbot-map](https://github.com/Robbyant/lingbot-map)
- Original Model Weights: [https://huggingface.co/robbyant/lingbot-map](https://huggingface.co/robbyant/lingbot-map)

---

## 📄 License

This model weight repository is released under the **Apache License 2.0**.