Hi3D open-sources v1.0 mesh refiner with up to 1536 resolution, built on Sparc3D
Chinese 3D generation company Hi3D has open-sourced Hi3D v1.0, a Sparc3D-based model that refines a coarse mesh and a reference image into high-resolution geometry, up to an effective 1536 cubed sparse grid. Inference code is on GitHub and model files are on Hugging Face. The license allows commercial use, with a separate license required above 200,000 monthly active users.
Chinese 3D generation company Hi3D has open-sourced Hi3D v1.0, a model that refines a coarse 3D mesh into high-resolution geometry at up to an effective 1536 cubed sparse extraction grid. Inference code is on GitHub, model files are on Hugging Face, and a free online demo is available, according to Pandaily.
What the model does
Hi3D v1.0 does not generate an object from scratch. It takes a coarse mesh and a reference image of the same object and outputs a more detailed mesh.
It is built on Sparc3D, a high-resolution 3D modeling framework presented at NeurIPS 2025 that focuses on representing and reconstructing complex geometry so thin parts and surface detail survive. Hi3D keeps that core architecture and expanded the training data to about two million curated 3D assets. The team says it validated this version in its own production workflow before releasing it.
How the resolution settings work
The pipeline samples sparse latents, decodes them with a self-pruning sparse VAE, extracts a mesh, and then repeats the process:
- 512: one pass
- 1024: two passes
- Top setting (1536): three passes, reaching an effective sparse extraction grid of 1536 cubed
The project notes that these figures describe nominal sparse grids, not a dense latent tensor or a fixed triangle count. Examples in the technical report include an armored vehicle whose angled armor, tire treads and underside stay consistent across front, back and bottom views, and a winged sculpture whose individual feathers remain distinct.
Hardware requirements
The requirements are steep:
- Linux and Python 3.10
- An NVIDIA GPU with compute capability 8.0 or newer
- No CPU path
The highest resolution setting has only been validated on a 96GB NVIDIA H20. Memory needs on smaller cards have not been measured. Training code and the paper's evaluation harness are not included in the release.
License terms
Hi3D v1.0 ships under the Hi3D Open License Agreement Version 1.0, which allows commercial use. Products or services that pass 200,000 monthly active users must notify Hi3D and apply for a separate license within 60 days. Original Apache-2.0 and MIT portions keep their own terms.
Relationship to Hi3D's commercial models
Hi3D's newer commercial models are not part of this release. The company describes v1.0 as an earlier, production-tested version that developers and businesses can build on. Its separate Hi3D Agent product links model generation to editing, rigging and 3D-printing preparation.
What this means
Hi3D v1.0 targets a specific stage of the 3D workflow: upgrading an existing rough mesh using an image reference, rather than creating assets from nothing. Teams that already have coarse geometry can test it through the online demo before committing hardware.
The practical barrier is compute. Only the top 1536 setting's validation on a 96GB H20 is documented, and requirements for smaller GPUs are unmeasured, so anyone planning to run the highest setting locally has to test their own hardware. The absence of training code and the evaluation harness means the results can be run but not reproduced from scratch.
The license permits commercial use but sets a 200,000 monthly active user threshold, above which a separate agreement with Hi3D is required. Because the newer commercial models are excluded, v1.0 is an earlier version rather than the company's current best.
Related Articles
Liquid AI releases LFM3: six 2B-class models on one backbone, led by a 1.7B text model with 256K context
Liquid AI has released LFM3, a family of six 2B-class models built from LFM3-1.7B-Base, covering text, vision, audio, omni and decision tasks. According to Liquid AI, LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters and offers a 256K-token context window.
SoMark open-sources OXR-1.0, a 1.3B document parser claiming first place on three benchmarks
SoMark has open-sourced OXR (Optical Everything Recognition), a 1.3B-parameter document parsing model. According to the team, it ranks first on OmniDocBench (97.18), olmOCR-Bench (87.84) and DocParsingBench (87.76), and parsed 100 test pages faster than PaddleOCR-VL-1.6 and MinerU2.5-Pro.
Tencent Hunyuan open-sources Hy-SimulMT, a 30B-A3B speech-to-speech interpreter with about 3 seconds of lag
Tencent Hunyuan has open-sourced Hy-SimulMT, a 30B-A3B mixture-of-experts model that translates Chinese and English speech in real time while cloning the speaker's voice. Tencent reports about 3 seconds of StreamLAAL lag on the long-form RealSI benchmark, the lowest among the systems it compared. The weights are listed under Apache 2.0.
EdgeRock-4B holds KV cache at 72 MiB up to 256K context, with 12,124 tokens/s on one 80GB GPU
EdgeRock-4B, an open-source model converted from Qwen3.5-4B, keeps its attention KV cache fixed at 72 MiB at 256K context, a 99.1% reduction versus the source model. On a single 80GB Hopper GPU it reaches 12,124 tokens/s at 512 concurrent sequences, which the developers report is 6.42x the source model. Weights and inference code are public, with commercial use allowed.
Comments
Loading...