Tencent Hunyuan 3D Lite version is here. It supports consumer-grade graphics cards. Is 3D modeling going mainstream for everyone?
On August 15th, Tencent's Hunyuan team launched the Lite version of the 3D world model. Compared with the previous requirement of 26GB of video memory, by introducing the dynamic FP8 (8-bit floating-point format) quantization technology this time, the video memory requirement is directly reduced to less than 17GB, allowing consumer-grade graphics cards to run smoothly.
Previously, although the FP32 version of Tencent's Hunyuan 3D world model could retain all details completely, it occupied an extremely high amount of video memory - its parameters might exceed one billion, and usually required a GPU with a large-capacity VRAM to improve the inference speed. Therefore, consumer-grade graphics cards simply could not support it.
To put it simply, FP32, FP16, and FP8 represent different "precision levels". In the past, when the high-precision FP32 technology was used, although extremely high precision restoration could be achieved, it would occupy a large amount of video memory and might also retain unnecessary details (for example, the sky texture of the background actually did not need to be so meticulously crafted).
And the core of this dynamic FP8 quantization technology lies in its ability to monitor the data distribution during the model operation in real-time and dynamically adapt to different modules: Most key areas use FP16 precision, while non-key parts such as the above-mentioned background textures are dynamically adjusted to FP8 precision.
This technology significantly reduces the video memory occupation. Although the precision is appropriately reduced in some areas, it allows individual players to easily use the 3D world model.
Tencent Hunyuan 3D Reinvents the 3D Modeling Process
Tencent Hunyuan 3D world model is the first open-source and editable world generation model in the industry. It can directly generate a complete, editable, and interactive world model based on the pictures or text information provided by users, and can be directly applied to scenarios such as game development, special effects production, and educational simulation.
Image source: Tencent Hunyuan 3D official website
Compared with the previous 3D model AI generation function of Tencent Hunyuan model, the content generated by the 3D world model launched this time is more abundant, covering multiple factors such as environmental style, indoor and outdoor scenes, and light rendering. Traditional 3D scene development takes an extremely long time. Just developing a main building scene may take several weeks or even longer. The efficiency improvement brought by this one-click generated scene completely exceeds users' imagination.
So, how does the Hunyuan 3D world model quickly generate a 360° immersive visual space in the face of such complex scene development?
From the model architecture of Hunyuan World Model 1.0, the panoramic world image generation technology, as a unified proxy system connecting text, pictures, and the world, first generates a panoramic view of the initialized world, thus achieving a 360° full-coverage scene.
Image source: Tencent Hunyuan 3D official website, Model architecture of Hunyuan World Model 1.0
Subsequently, the system decomposes the entire 3D world into different clear levels, such as foreground and background, ocean and ground, ground and sky, etc., and then reconstructs the 3D world based on these levels, finally forming a 3D world model.
Compared with the traditional 3D scene development where every detail needs to be meticulously crafted, consuming a large amount of time and human resources, this one-click generated scene can not only save a lot of time but also output standardized and navigable 3D Mesh assets, which are compatible with tools such as Unity and Unreal Engine.
Moreover, the precision of the generated content has reached a level where it can be directly used: the details in the attention area of the foreground are well presented, the separation between the background and the foreground is sufficient, and there are no problems such as unclear boundaries and blurred light and shadow.
However, after experiencing the Hunyuan 3D world model on the official website, it can be found that it cannot fully restore all the requirements in the text, and can only restore the general scene requirements, light and shadow colors, and details in the foreground area.
For example, the text requirements corresponding to the following picture mention elements such as a mechanical world and robots, but these are not presented in the generated scene. The system only extracts the words related to constructing the general world scene, such as the cyberpunk wasteland style and the red setting sun in the sky, and then separates the foreground and the background - deconstructs the "abandoned amusement park" as the foreground content and the red setting sun as the background sky content, and then reconstructs the 3D world scene based on these levels. That is to say, it only restores the general scene requirements.
Image source: Tencent Hunyuan 3D official website
It can be clearly seen that the Hunyuan 3D world model currently cannot meet users' personalized needs. However, it can already initially construct the foreground, background, and simple scene details, which can save a lot of time in work such as game development.
In addition, the 3D world model generated according to users' requirements is also highly playable for ordinary players. The direct output of 3D Mesh assets brings format unity and reduces the learning cost. When AI can complete the scene deconstruction and 3D construction work, users' subjective initiative becomes the only variable determining the generated scene.
Is the 3D Model Going Mainstream in 2025?
Tencent's purpose of popularizing the Hunyuan 3D world model to consumer-grade graphics cards this time is very clear - to attract a large number of developers and creators to join the "Tencent Hunyuan 3D" ecosystem. This model supports the full-process content generation from 3D models to 3D world scenes, allowing users to create their own virtual worlds.
Currently, there are many large AI models supporting 3D model generation on the market, such as Tripo AI, Meshy AI, GENIE, etc. However, the fact that many players are flocking to the 3D track has led to a high degree of homogenization of product functions, which also reflects from the side that "bringing real-world scenes into the virtual world" has become the core function that all manufacturers must compete for.
Among these AI tools, Tripo AI, an AI 3D foundation model released by the Silicon Valley startup VAST in 2024, stands out with its unique product structure.
Different from Tencent Hunyuan 3D, which is targeted at a wider range of users, Tripo AI is more oriented towards professional creators: After entering the page, users can directly generate 3D models through text or pictures, and the adjustable parameters are relatively rich - it not only supports the texture generation function that all current mainstream AI 3D models have but also can automatically split model parts, allowing each disassembled part to be edited separately; it even supports binding basic animations to model parts and demonstrating them, although there may be occasional problems with part deformation during the demonstration. Overall, Tripo AI is a mature AI 3D tool that can be adapted to multiple scenarios.
Image source: Tripo AI
Meshy AI, also launched in 2024 (created by a domestic team), although it also supports directly generating 3D models through text and images, its core advantage lies in its more complete community function: users can browse the 3D model works of other creators in the community, and the platform has a clear and detailed classification of models, and also marks key information such as interaction volume, likes, and whether 3D printing is supported. This design allows novice users to directly download ready-made 3D models for use, and also improves the dissemination and activity of the community.
Image source: Meshy AI
The GENIE tool launched by Luma AI, in addition to supporting text-to-3D model conversion and multi-format (such as OBJ, FBX, etc.) export to adapt to different scenarios, its biggest highlight is that it provides an API interface - users can directly convert video content into 3D models through this interface, forming a differentiated competitive advantage.
It is not difficult to see that the above products have all broken through the homogenized competition with their own characteristics, and Tencent Hunyuan 3D is no exception. Although there is no significant gap between its 3D model generation function and other tools, "high free quota" is its core advantage: on the Hunyuan AI 3D official website, each user can generate models for free 20 times a day, and can regain the quota by sharing with friends after the quota is used up. This promotion strategy of "exchanging quantity for users" has been quite successful. Before the release of the Lite version of the 3D world model, the download volume of its community models had reached 2.3 million times, making it one of the most popular 3D open-source model platforms in the world.
Image source: Tencent Hunyuan AI
Tencent's launch of the Lite version of the Hunyuan 3D world model compatible with consumer-grade graphics cards this time will undoubtedly attract more creators to join its ecosystem. The growth of the user scale will further promote feedback iteration and application scenario expansion: taking the currently popular VR glasses as an example, the 3D world model files exported by Hunyuan 3D can be directly imported and used. Users only need to have VR equipment to immerse themselves in the virtual scenes they created anytime, anywhere, realizing the linkage between the ecosystem and hardware; at the same time, the AI 3D foundation model enables ordinary users to easily create highly customized 3D models, forming a synergy with 3D printers.
More importantly, the characteristic of "almost zero learning cost" of AI 3D is driving its rapid penetration into various industries: in scenarios such as architectural planning, interior design, and e-commerce display, 3D visual content is easier to understand than text or traditional drawings. Staff can output scene content without complex learning, greatly reducing the time for repeated modeling; this linkage between "virtual models + real industries" can not only improve user stickiness but also make users feel a sense of belonging through highly customized content - all these trends indicate that 3D models are bound to become popular in 2025.
Xiaolei believes that future AI 3D models will further integrate professional scene models and creative styles, attract more vertical users through segmented fields and usage scenarios, continuously expand the ecological boundaries, and penetrate into various daily life scenarios.
And this is precisely the core significance of this wave of 3D model popularization - in the current era of the integration of reality and virtuality, enabling everyone to have the ability to build a 3D virtual world.
Will 3D Modelers Lose Their Jobs Due to the Popularization of 3D Models?
However, there has always been an argument online that with the popularization of 3D models, 3D modelers will face the risk of unemployment. Xiaolei does not agree with this view.
It is undeniable that these tools capable of quickly generating 3D models will inevitably have an impact on the industry. The "fast and efficient" advantage of AI models is indeed difficult for humans to match; but as mentioned above, current AI 3D models still cannot achieve true user personalization - the products they generate are essentially "replicated content" based on the learning data of large models.
And this kind of content lacking in personality will ultimately not become excellent works. Whether it is game modeling or architectural design, what really makes people remember are always those ingenious designs: the details carefully polished by 3D modelers and the ingenuity carefully considered to meet users' needs. Therefore, Xiaolei believes that with the current capabilities of AI 3D models, it is basically impossible to completely replace 3D modelers; on the contrary, as a tool capable of efficiently executing repetitive instructions, it is more suitable to be an "assistant" for modelers to improve efficiency.
Actually, if you think about it, this "AI-assisted creation" model has long penetrated into various industries. However, limited by the problem of content homogenization, AI often can only stay in the "repetitive basic construction" stage.
This is also the reason why Lei Technology still adheres to original content creation today when AI writing tools are becoming more and more convenient and popular. Xiaolei always believes that truly in-depth and warm good articles will never be overshadowed by the existence of AI.
This article is from "Lei Technology" and is published by 36Kr with authorization.