# Where to Find LingBot-Map Model Checkpoints and Variants: Complete Download Guide

> Download LingBot-Map model checkpoints and variants from Hugging Face and ModelScope. Access PyTorch versions lingbot-map long stage1 and more, all in one place.

- Repository: [Robbyant/lingbot-map](https://github.com/Robbyant/lingbot-map)
- Tags: getting-started
- Published: 2026-07-06

---

**LingBot-Map model checkpoints are publicly hosted on Hugging Face and ModelScope under the `robbyant/lingbot-map` repository, with three PyTorch variants available: `lingbot-map-long.pt`, `lingbot-map.pt`, and `lingbot-map-stage1.pt`.**

The Robbyant/lingbot-map repository distributes its pretrained weights as standard PyTorch `.pt` files through major model-hosting platforms. These checkpoints capture different training stages and optimization targets, allowing you to select the appropriate variant for your specific video understanding workload.

## Available LingBot-Map Model Variants

According to [`README.md`](https://github.com/Robbyant/lingbot-map/blob/main/README.md) lines 133-140, the repository provides three distinct checkpoint variants:

- **lingbot-map-long** – Optimized for very long video streams and large-scale scenes.
- **lingbot-map** (default) – The balanced checkpoint used in the paper and most demos.
- **lingbot-map-stage1** – The stage-1 training checkpoint; can be loaded into the VGGT backbone for bidirectional inference (c2w).

All variants are stored as PyTorch tensor files and accessible through both Hugging Face and ModelScope repositories.

## Downloading LingBot-Map Checkpoints

### Using the huggingface_hub Python Library

Programmatically fetch specific checkpoint files using the `huggingface_hub` library. This method handles caching and returns the local path:

```python
from huggingface_hub import hf_hub_download

# Choose the variant you need

repo_id = "robbyant/lingbot-map"
filename = "lingbot-map-long.pt"          # or "lingbot-map.pt", "lingbot-map-stage1.pt"

# Downloads the file into a local cache and returns the path

ckpt_path = hf_hub_download(repo_id=repo_id, filename=filename)

print(f"Checkpoint saved at: {ckpt_path}")

```

### Using the Hugging Face CLI

Download checkpoints directly from the terminal using the Hugging Face CLI tool:

```bash

# Download the long-form variant

huggingface-cli download robbyant/lingbot-map lingbot-map-long.pt

# Download the default variant

huggingface-cli download robbyant/lingbot-map lingbot-map.pt

```

### ModelScope Alternative

For users preferring ModelScope, identical files are hosted at `https://www.modelscope.cn/models/Robbyant/lingbot-map`. You can download the same `.pt` files through the ModelScope web interface or their Python SDK.

## Loading Checkpoints in Your Code

The LingBot-Map codebase implements checkpoint loading using standard PyTorch deserialization. In [`demo.py`](https://github.com/Robbyant/lingbot-map/blob/main/demo.py) at lines 154-155, the inference script loads weights via `torch.load`:

```python
import torch

ckpt_path = "lingbot-map-long.pt"          # path obtained from the download step above

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
ckpt = torch.load(ckpt_path, map_location=device, weights_only=False)

# The checkpoint can now be passed to the model constructor

# model = LingBotMapModel.from_checkpoint(ckpt)

```

The base aggregator class in [`lingbot_map/aggregator/base.py`](https://github.com/Robbyant/lingbot-map/blob/main/lingbot_map/aggregator/base.py) at line 221 contains additional loading logic for library usage, while [`benchmark/methods/lingbot_map.py`](https://github.com/Robbyant/lingbot-map/blob/main/benchmark/methods/lingbot_map.py) at line 130 handles checkpoint initialization for evaluation scripts.

## Summary

- **Three variants** exist: `lingbot-map-long` (extended video), `lingbot-map` (default), and `lingbot-map-stage1` (bidirectional inference).
- **Hosting platforms**: Hugging Face (`robbyant/lingbot-map`) and ModelScope (`Robbyant/lingbot-map`).
- **Download methods**: `huggingface_hub` library, `huggingface-cli`, or direct web UI.
- **File format**: PyTorch `.pt` files loaded via `torch.load()`.
- **Key implementation files**: [`demo.py`](https://github.com/Robbyant/lingbot-map/blob/main/demo.py) (lines 154-155), [`lingbot_map/aggregator/base.py`](https://github.com/Robbyant/lingbot-map/blob/main/lingbot_map/aggregator/base.py) (line 221), and [`benchmark/methods/lingbot_map.py`](https://github.com/Robbyant/lingbot-map/blob/main/benchmark/methods/lingbot_map.py) (line 130).

## Frequently Asked Questions

### What are the differences between the three LingBot-Map checkpoint variants?

The **lingbot-map-long** variant is optimized for very long video streams and large-scale scenes requiring extended context windows. The default **lingbot-map** checkpoint provides balanced performance characteristics used in the primary paper experiments. **lingbot-map-stage1** represents an intermediate training checkpoint specifically compatible with the VGGT backbone for bidirectional camera-to-world (c2w) inference tasks.

### How do I download LingBot-Map checkpoints programmatically?

Use the `huggingface_hub` Python library with `hf_hub_download(repo_id="robbyant/lingbot-map", filename="lingbot-map.pt")` to fetch files with automatic caching. Alternatively, install the Hugging Face CLI and run `huggingface-cli download robbyant/lingbot-map lingbot-map.pt` to retrieve weights directly to your local filesystem.

### Can I use ModelScope instead of Hugging Face to download the weights?

Yes, identical checkpoints are mirrored on ModelScope at `https://www.modelscope.cn/models/Robbyant/lingbot-map`. This provides an alternative download source for users who prefer ModelScope's infrastructure or require access through Chinese cloud networks.

### Where is the checkpoint loading logic implemented in the source code?

The primary loading mechanism resides in [`demo.py`](https://github.com/Robbyant/lingbot-map/blob/main/demo.py) at lines 154-155, which calls `torch.load()` with `map_location` set to the available CUDA device or CPU. Additional loading implementations appear in [`lingbot_map/aggregator/base.py`](https://github.com/Robbyant/lingbot-map/blob/main/lingbot_map/aggregator/base.py) at line 221 for the base aggregator class, and in [`benchmark/methods/lingbot_map.py`](https://github.com/Robbyant/lingbot-map/blob/main/benchmark/methods/lingbot_map.py) at line 130 for benchmark evaluation scenarios.