SpatialLM1.1-Qwen-0.5B

SpatialLM1.1-Qwen-0.5B

Reads 3D point clouds and writes a structured scene layout with walls, doors, windows and labelled object boxes. For researchers and developers working on robotics, navigation and 3D scene understanding.

Made by manycore-research on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: SpatialLM1.1-Qwen-0.5B is a small language model that reads 3D point cloud data and writes out a structured description of the scene. It labels walls, doors and windows, and it draws oriented boxes around objects with their categories. It is built on Qwen2.5-0.5B-Instruct and has about 604 million parameters.

Who it suits: It suits researchers and developers who work with scanned indoor spaces and need machine-readable layouts rather than raw dots.

Industry | Who uses it | How they use it | Example
Robotics | Robot developers | Give a robot a structured view of a room | Finding doors and walls from a scan
Navigation | Autonomy engineers | Feed scene layouts into planning code | Listing objects with their boxes and categories
Research | 3D vision researchers | Study scene understanding from point clouds | Testing on video or LiDAR captures

You can skip it if you need a commercial license, since this one allows non-commercial use only, or if you work only with text and images.

Where to use it: It runs in Python with the transformers library, on Windows, macOS or Linux. The card says it handles point clouds from several kinds of capture.

• Python with the transformers library
• Point clouds made from monocular video sequences
• Point clouds from RGBD images
• Point clouds from LiDAR sensors

Use case | Who | What they get back
Room layout from a scan | Robotics teams | Walls, doors and windows as structured output
Object detection in 3D | Vision researchers | Oriented boxes with a semantic category each
Custom category search | Developers | Detection of categories you specify

How to set it up: You need Python and a way to turn your capture into a point cloud. The weights are about 2.4 GB in F32 according to Hugging Face. The model card and the project repository hold the run steps for the full pipeline.

• Python
• About 2.4 GB of disk for the F32 weights
• A point cloud of your scene

1. Install the Hugging Face download tool.
2. Download the model files.
3. Follow the model card to prepare a point cloud and run the model.

pip install -U huggingface_hub
hf download manycore-research/SpatialLM1.1-Qwen-0.5B

How to use it: You give the model a point cloud of a space and it answers with a structured list of the architectural elements and objects it finds. Version 1.1 also accepts a list of categories you want detected. The exact run commands sit in the full model card.

Tool | What it does
hf download | Fetches the model files to your computer
transformers | Loads the model for use in Python

Example: You scan a living room, turn the capture into a point cloud, and run the model on it. It returns the layout as text, with the walls, doors and windows and then each object as an oriented box with a category such as sofa or table.

For newcomers and experts: If you are new, think of it as a model that describes a 3D scan in words and numbers. Download the files first and read the card to see how a scan becomes a point cloud.

If you know the field, version 1.1 doubles the point cloud resolution and uses a Sonata point cloud encoder feeding a Qwen language model. It adds detection of user-given categories. The authors report acceptance at NeurIPS 2025.

Capture from video, RGBD or LiDAR
Convert to a 3D point cloud
SpatialLM encodes the points
Language model writes the scene layout

Good to know: This 0.5B model is the smaller of the two 1.1 releases, next to a 1B Llama-based one.

• License: CC BY-NC 4.0, so non-commercial use only
• Cost to run: no API key, account or paid service is named
• Data it reaches: the point clouds you feed it, plus files downloaded from Hugging Face
• Not for: commercial products under this license, or tasks that do not involve 3D scenes

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 604M
• Weights: about 2.4 GB in F32; about 302 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,054
• Likes: 43
• License: cc-by-nc-4.0
• Based on: Qwen/Qwen2.5-0.5B-Instruct
• Fine-tunes: 3
• Tags: transformers, safetensors, spatiallm_qwen, text-generation, conversational

Links

• Model: https://huggingface.co/manycore-research/SpatialLM1.1-Qwen-0.5B
• Files: https://huggingface.co/manycore-research/SpatialLM1.1-Qwen-0.5B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download manycore-research/SpatialLM1.1-Qwen-0.5B

Then follow the usage in its model card. Nothing in this kit runs by itself.