Alibaba's AutoNavi launches world model studio running on single RTX 5090

Alibaba Group's AutoNavi unit launched ABot-World Studio, a general world model that unifies interactive video generation with 3D scene construction on a single platform, running stable inference for more than one hour on one Nvidia RTX 5090 GPU.

The product, now open for testing, lets users generate interactive AI worlds from text prompts or images and save outputs as video and 3DGS files. It is the first platform to combine two previously separate technical approaches — interactive video generation, which prioritizes immersion but ends once the video is generated, and 3D scene generation, which builds spatial assets but lacks continuity between worlds.

ABot-World Studio solves the long-term instability problem that has plagued video generation models. In official tests, continuous inference ran for more than one hour without crashes or quality degradation, far exceeding the roughly one-minute generation limit of comparable world models. The system maintains video quality equivalent to the first frame throughout the session, according to AutoNavi's published specifications.

The studio is powered by two underlying models: ABot-World0 for video generation and ABot-3DWorld0 for 3D generation. ABot-World0 integrates scene navigation with character control into a unified architecture, using algorithmic optimizations to reduce error accumulation during extended inference. ABot-3DWorld0 shares its inference pipeline with the previously released ABot-Earth0.5, achieving what the company describes as the first unified architecture spanning indoor, street-level, and urban aerial spatial generation at full scale. Both models can perform inference locally on a single consumer-grade GPU without requiring dedicated computing infrastructure — a rarity among current world models that typically demand server-class hardware.

The studio introduces a "Spacetime Portal" mechanism that allows users to traverse between different 3D worlds. A wooden door in a Jiangnan water town scene can link to a cybercity, or a ferry landing at Fenglingdu can connect directly to a mountain pass, weaving isolated scenes into an interconnected exploration network. All generated 3D worlds can be saved and shared, forming what AutoNavi describes as an open public repository.

The ABot-World series models are fully open source and have been published on GitHub, Hugging Face, and Reactor. The company said the technology has applications in embodied intelligence training for robotics, film and video storyboarding — where it can reduce creative validation cycles from weeks to hours — and tourism, culture, and education, where users can experience historical scenes as participants rather than passive viewers.

The launch deepens Alibaba's commitment to AI-native products and positions AutoNavi, best known for its mapping and navigation services, as a player in the generative world model space. The ability to run on consumer hardware — a single RTX 5090 — lowers the barrier to entry for developers and researchers who cannot access the server clusters typically required for comparable systems. Nvidia's RTX 5090, the company's latest consumer flagship GPU, provides the memory bandwidth and compute capacity needed for real-time inference at the edge.

For investors, the question is whether open-source world models running on consumer hardware can erode the competitive moat of proprietary systems that require cloud-scale infrastructure. Alibaba's cloud division, which competes with Amazon Web Services and Microsoft Azure in Asia, could benefit from developer adoption of the ABot-World ecosystem. BABA-W shares traded up 0.09% on the Hong Kong exchange on the day of the announcement, though the stock's 22.6% short-selling ratio suggests the market has yet to price in the product's potential impact.

This article is for informational purposes only and does not constitute investment advice.