Hi3D open-sources the commercially usable 3D model V1.0 and launches a 3D creation Agent

Domestic AI 3D large model Hi3D has arrived! Open source and commercially usable, taking on and beating GPT-6

In terms of 3D generation results, it is indeed still the case that specialized models do specialized work.

From left to right: the original image, the GPT-6 Astra model generation, and the Hi3D Agent model generation.

Astra, why are you the ugliest!

It can be seen that Astra built a recognizable basic form, but the results generated by Hi3D Agent and its V3.0 model are closer to the reference image in outline, volume relationships, and local details.

Hi3D Agent is an AI assistant launched by Hi3D for 3D creators; behind it, the latest model Hi3D V3.0 is the industry's first commercially usable 2048³ ultra-high-precision domestically developed AI 3D model.

This also reflects the division of labor between the two types of models: large language models are responsible for understanding creative intent and organizing tasks, while specialized 3D models are responsible for generating three-dimensional results. Only by combining the two can the Agent translate users' requirements into concrete 3D works.

Today, Hi3D also officially opened the V1.0 model code. This is a high-resolution geometry refinement model built on Sparc3D, and it is also an earlier completed version in the Hi3D model series that has been validated through production workflows.

According to the team, V1.0 was trained on about 2 million filtered 3D assets and underwent validation in the team's production workflow before release.

The team also hopes to use this stable, production-validated version as a foundation that developers and enterprises can continue to research, deploy, and develop on.

Generation, modification, application: Hi3D takes over the creative workflow

The technical origins of Hi3D V1.0, open-sourced today, can be traced back to Sparc3D, made public in 2025.

This work was published at NeurIPS 2025 and proposed a high-resolution 3D modeling framework focused on addressing the representation and reconstruction of complex geometric structures, allowing small parts and surface details to be more fully preserved.

Its generation results be like:

English text is being prepared.

It is still quite impressive, and one can clearly see Sparc3D's ability to handle complex geometric details.

Hi3D V1.0 retains this core architecture and expands the training data scale to about 2 million filtered 3D assets, further enhancing high-resolution geometry generation capability.

Let's take a specific look at V1.0's creative results:

The armored vehicle shown in the V1.0 technical report demonstrates the model's complete generation capability for three-dimensional objects.

The front clearly shows the vehicle's sloped armor surfaces and tire treads; turning to the back and bottom, the body structure remains coherent, and the segmentation of the rear panels and the positional relationship between the chassis and wheels can also be discerned.

For areas not shown in the reference image, the model made reasonable completions.

Now let's look at more complex and small structures. For example, this sculpture wearing armor, holding a greatsword, and with wings spread behind its back:

After upscaling with V1.0, the wings behind the figure are individually distinct, and one can even see the layers formed by overlapping feathers.

From the overall structure of the vehicle body to the tiny gaps between feather pieces, these two cases present V1.0's geometric performance at different scales.

After the model generates results, creators can also continue modifying them in Hi3D Agent.

Hi3D Agent is a 3D creation Agent built on Hi3D's professional AI 3D models as its foundation, connecting intent understanding, task planning, model generation, continuous modification, engineering optimization, and application delivery.

For example, the user points out issues with floating parts and connection positions, asks to adjust the steering wheel, and clearly states: next, they want to print the model.

Following these requirements, the Agent lists the subsequent processing steps and continues adjusting the shape, local structure, geometry, texture, and mesh around the existing model, bringing the result closer to the creative goal.

In game and interactive content, the Agent can continue adding characters, props, and scenes, bringing generated assets into interactive environments and advancing them to playable demos.

English text is being prepared.

For animation and real-time applications, Hi3D Agent can perform lightweighting, skeletal rigging, motion, and facial expression processing on characters, enabling static models to enter animation or real-time content workflows.

English text is being prepared.

In physical creation, Hi3D Agent connects with the Hi3D platform's part splitting, connection structures, color schemes, and print preparation capabilities, continuing to transform digital models into physical works such as figures and designer toys.

English text is being prepared.

At the same time, generated results can also be downloaded and brought into Blender, game engines, or other professional software for further processing.

From model generation to modification and refinement, and then to subsequent applications, Hi3D can support creators in completing work at different stages, gradually turning ideas into usable works.

For the open-sourced V1.0, how did the team improve the model's structural integrity and detail performance? The technical report also provides more specific training and evaluation information.

Expanding training data and refining geometry: the technical upgrades of Hi3D V1.0

Hi3D V1.0 retains the Sparc3D architecture, with a focus on improving data scale and high-resolution capability.

This version retains the original core architecture, with the main work focused on expanding the training data, high-resolution training and output, and validation of the production pipeline.

First, the expansion of training data. Compared with the training setup of 500,000 assets in the Sparc3D paper, Hi3D V1.0 expanded the training scale to about 2 million curated 3D assets, four times the original.

This is related to the characteristics of the single-image-to-3D task. A single reference image can only provide partial information about an object; the unseen back and occluded regions still require the model to make reasonable completions based on the shape patterns it has learned.

Richer training assets allow the model to encounter more types of objects and structures, providing more sufficient learning samples for these inferences.

So, with the training data in place, how does the model gradually render details?

Combining the architecture diagram in the report, this process can be understood as: first determine the overall structure, then refine it level by level with reference to the image.

The upstream model first generates a coarse mesh from the image, providing the object's overall outline and main parts. Then, V1.0 references both the image and the coarse mesh to perform sparse geometric refinement around the object's surface.

Each time a level is completed, the system extracts a 3D mesh. This result can be output directly or used as the basis for the next round of refinement, gradually progressing to the 1024³ and 1536³ levels.

1536³ refers to the resolution of the geometry extraction mesh. It can be understood as the spatial scale used to describe the object's shape: the finer the scale, the more capable it is of expressing small undulations and gaps. The system adopts a sparse approach, mainly processing the effective regions near the surface.

The feather case shown above is able to achieve distinct individual strands precisely because of this resolution.

Currently, the latest Hi3D V3.0 has raised the geometric resolution to

Beyond resolution, there are also geometric details and downstream usage performance.

According to its technical report, Hi3D V1.0 achieves multi-view fidelity and detail richness of 0.75 and 0.841 respectively, ranking first among the comparison methods listed in the report.

At the same time, its slicing success rate reaches 90%, demonstrating its usability for entering the subsequent 3D printing preparation process.

△From the V1.0 technical report

Slicing is the preparation step that converts a 3D model into layer-by-layer instructions for a printer. This result reflects the performance of the generated mesh entering the print preparation process; whether it can actually be printed smoothly also depends on conditions such as wall thickness, supports, and materials.

According to the team, before its release, V1.0 had already entered real user-facing production workflows for testing and use. In addition to observing generation results, the team also tested in actual creative tasks whether these assets could continue to be processed and enter subsequent applications.

Open model and tool interfaces let developers and enterprises build 3D applications

The Hi3D V1.0 released this time is now officially available on GitHub and Hugging Face.

The official release has disclosed the system architecture and effect comparisons, and opened the source code; the project page also has an online Demo where you can experience the related generation capabilities.

This model is mainly aimed at researchers, developers, and enterprises.

Researchers can use it as a foundation for technical validation and model research, exploring complex structure generation and detail preservation;

Developers can integrate the geometric refinement capability into professional tools to build vertical applications and their own workflows;

Enterprises can, in accordance with the officially released licensing and deployment terms, explore private deployment, internal system integration, and business adaptation.

Beyond the open model, Hi3D can also connect to external applications via MCP.

Developers can call Hi3D in Agents or tools that support MCP to complete operations such as submitting generation tasks, querying progress, and obtaining model results, and then integrate these steps into their own business processes.

Hi3D Agent, meanwhile, directly serves creators. Users can describe their needs in natural language and continue to make modification requests around the generated results, letting the Agent assist in advancing the creation.

The three have their own roles: the open model provides a foundation for independent development, MCP lets external applications call platform capabilities, and the Agent helps users directly complete creation.

This also corresponds to Hi3D's positioning as AI 3D infrastructure: providing support for model development, application integration, and actual creation.

These efforts also jointly point toward the "Creator First" direction proposed by Hi3D:

Helping beginners enter 3D creation more easily, reducing repetitive operations for professional creators, and leaving more time for design and modification.

As domestic AI 3D technology continues to mature, the distance between imagination and works is also shrinking bit by bit. Ordinary people, starting from a single image or an idea, can also create their own 3D works.