SmolRGPT: An Efficient Multimodal Spatial Reasoning Model with Only 600 Million Parameters

SmolRGPT, with a compact scale of 600 million parameters, has achieved breakthrough performance in multimodal spatial reasoning tasks and won third place in the AI City Challenge 2025, demonstrating the huge potential of small models in specific domains. At the same time, it provides important insights for solving practical problems such as high computing costs, deployment difficulties, and high energy consumption of large models.

Currently, the mainstream in the AI field focuses on the parameter scale race, but this brings problems such as high computing costs, deployment difficulties, and huge energy consumption. Spatial reasoning is a key capability of multimodal AI, referring to the understanding of object spatial positions, relationships, motion trajectories, etc. Its application scenarios include autonomous driving, robot navigation, video surveillance, augmented reality, etc., which require the model to understand 3D spatial structures and dynamic changes.

The key to SmolRGPT's powerful spatial reasoning with limited parameters:

1. Domain-Specific Design: Focus on spatial reasoning, concentrating parameters on relevant representation learning;
2. Efficient Visual Encoding: A lightweight encoder extracts spatial features, optimizing the capture of position, distance, and direction;
3. Multimodal Fusion: Efficient fusion of visual features and text queries, supporting natural language spatial question answering;
4. Reasoning Capability Enhancement: Possesses multi-step reasoning ability to handle complex spatial queries.

The AI City Challenge is a top event in the intelligent transportation field, attracting top teams worldwide in 2025. SmolRGPT won third place with significant advantages:

• Parameter Efficiency: Its scale is an order of magnitude smaller than most participating large models;
• Reasoning Speed: Lightweight design brings speed advantages in deployment scenarios;
• Resource-Friendly: Can run on consumer-grade GPUs or even edge devices. This proves that model quality is more important than scale in specific domains.

Its lightweight features give it unique advantages in multiple scenarios:

• Edge Deployment: Run in real time on edge nodes such as cameras and in-vehicle devices, reducing cloud latency;
• Cost-Sensitive Scenarios: Reduce hardware and operational costs for security surveillance and smart city projects;
• Real-Time Interaction: Low latency is suitable for applications requiring instant feedback;
• Research and Education: Easy to understand and modify, lowering the threshold for multimodal AI research.

SmolRGPT has been open-sourced on GitHub, providing weights, code, and examples to promote academic expansion, industrial integration, and technology popularization. Its success provides the following insights:

• Efficiency and performance can coexist: Task focus + architectural innovation allow small models to be comparable to large models;
• Application-driven development is more sustainable: Design models based on actual needs rather than blindly pursuing scale;
• Multimodal democratization: Lightweight models enable more developers to use advanced AI technologies.

SmolRGPT represents the direction of extreme efficiency in AI models. With the growth of edge AI demand, such dedicated models will become more important. For developers, it provides an example of achieving excellent capabilities with limited resources, and the 'small but refined' approach may be one of the key paths to the inclusiveness of AI technology.