SpatialLM
Manycore Tech (群核科技) of Hangzhou, behind Kujiale and Coohom, open-sourced its indoor spatial stack. SpatialLM turns point clouds into structured layouts, SpatialGen generates scenes.
Overview
Kujiale is the interior design software Chinese decorators and furniture retailers actually work in, and Coohom is its international version. What that gave Manycore Tech, founded in Hangzhou in 2011, is an enormous corpus of structured indoor scenes: rooms with labelled walls, openings and furniture placements. SpatialLM is that corpus turned into a model. Feed it a point cloud reconstructed from monocular video, an RGBD capture or a LiDAR scan, and it returns walls, doors, windows and oriented object bounding boxes as structured text that an application can parse directly, rather than a mesh you then have to interpret.
The open releases are deliberately small. SpatialLM1.1 ships as a 0.5B model built on Qwen-2.5 under Apache 2.0 and a 1B model built on Llama-3.2 under Meta's licence, both published with a test set and reproducible figures: 94.3 percent F1 at 0.25 IoU for layout estimation on Structured3D and 65.6 percent at the same threshold for object detection on ScanNet. The 1.1 update doubled point cloud resolution, moved to the Sonata encoder and added user-specified object categories. SpatialGen 1.0 works in the other direction, taking a 3D semantic layout and generating multi-view RGB, normal, depth and semantic maps, with training code released in January 2026 and dataset access opened by application in April 2026. Both sit on Hugging Face, GitHub and ModelScope, so access is not gated behind a mainland account.
Two things decide whether you can actually use this. The first is the licence patchwork, which is genuinely mixed rather than uniformly open: Apache 2.0 on the Qwen path, Meta's terms on the Llama path, CC-BY-NC-4.0 on one point cloud encoder and CreativeML Open RAIL++-M inherited from Stable Diffusion 2.1 on SpatialGen, plus a non-commercial FLUX.1 LoRA on the wireframe route. Any commercial deployment needs a component-by-component read. The second is that SpatialLM 1.5, announced on 27 August 2025 alongside SpatialGen with a phased open-source roadmap, still has not fully landed. Manycore listed on the Hong Kong exchange on 17 April 2026 as 00068.HK and closed its first day 144 percent above the offer price, which at least means the open line now has a public balance sheet behind it.
Key Features
- ✓ SpatialLM1.1 in two sizes: a 0.5B model on Qwen-2.5 under Apache 2.0 and a 1B model on Llama-3.2 under Meta's licence
- ✓ Accepts point clouds derived from monocular video, RGBD capture or LiDAR and emits walls, doors, windows and oriented object bounding boxes as structured text
- ✓ Published benchmarks with a reproducible test set: 94.3 percent F1 at 0.25 IoU for layout estimation on Structured3D and 65.6 percent F1 at 0.25 IoU for 3D object detection on ScanNet, both for the Qwen build
- ✓ SpatialGen 1.0 for layout-guided indoor scene generation, outputting multi-view RGB, normal, depth and semantic maps, with training code released January 2026 and dataset access opened April 2026
- ✓ User-specified object categories and a doubled point cloud resolution added in the 1.1 release, using the Sonata encoder
- ✓ Small enough to fine-tune on a single workstation GPU, tested on Python 3.11 with PyTorch 2.4.1 and CUDA 12.4
Where it holds
- • A 0.5B model that outputs parseable structured scene descriptions is genuinely deployable on a robot or an edge box, which almost nothing else in spatial AI manages
- • Numbers are published in the repository with a released test set rather than asserted in a press release, so you can reproduce them before committing
- • The training data advantage is real and hard to copy: Kujiale is the interior design tool Chinese decorators actually use, which gave Manycore a corpus of structured indoor scenes nobody else has
- • Weights are mirrored on Hugging Face, GitHub and ModelScope, so access works cleanly from either side of the firewall
Where it breaks
- • The licensing is a patchwork you have to read line by line. One point cloud encoder is CC-BY-NC-4.0, which quietly blocks commercial use of that path, and the Llama build carries Meta's terms rather than an open licence
- • SpatialGen 1.0 derives from Stable Diffusion 2.1 under CreativeML Open RAIL++-M and its FLUX.1 wireframe LoRA is non-commercial, so the generation half is not clean for a shipped product
- • SpatialLM 1.5 was announced in August 2025 with a phased open-source roadmap and the full weights are still unpublished, so the flagship stays internal while 1.1 is what you can actually run
- • Indoor only, and furniture-shaped. Walls, doors, windows and household objects are the vocabulary. It has no useful notion of mechanical assemblies, terrain or outdoor structure
- • English documentation is thinner than the Chinese material, and the SpatialVerse commercial platform is documented mainly for mainland customers
My Take
Robotics teams have a data problem more than a model problem, and this points straight at it. The 0.5B Qwen build takes a point cloud from a phone video and returns walls, doors and oriented object boxes as structured text, scoring 94.3 percent F1 at 0.25 IoU on Structured3D layout, and it is small enough to fine-tune on one workstation card. Read the licence table before you ship anything: the Qwen path is Apache 2.0, the Llama path carries Meta's terms and one point cloud encoder is CC-BY-NC-4.0, which is easy to miss and expensive to discover late.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Mixed: Apache 2.0 on SpatialLM1.1-Qwen-0.5B, the Llama 3.2 licence on SpatialLM1.1-Llama-1B, CC-BY-NC-4.0 on one point cloud encoder, CreativeML Open RAIL++-M on SpatialGen-1.0 with a non-commercial FLUX.1 wireframe LoRA
- Starting at:
- Free. Weights and code are published on Hugging Face, GitHub and ModelScope. What you may do with them depends on which artifact you take: Apache 2.0 on the Qwen build, Meta's Llama 3.2 licence on the Llama build, CC-BY-NC-4.0 on one of the point cloud encoders and CreativeML Open RAIL++-M on SpatialGen-1.0. The commercial products, Kujiale, Coohom and the SpatialVerse platform, are quoted separately and are not covered by these licences.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where SpatialLM holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to SpatialLM
Postshot
freemiumDesktop Gaussian splatting trainer from Jawset. Video or photo sets into radiance fields on a local NVIDIA GPU, with PLY and SPZ export. Windows only, subscription.
RoboBrain 2.5
open-sourceBAAI's embodied brain model (悟界·RoboBrain), Apache 2.0 at 4B and 8B, shipped in parallel NVIDIA and Moore Threads builds. It plans and points in 3D for a controller underneath.