Multimodal models get batch feature updates, spatial understanding becomes new focus
In November 2026, the AI industry saw a wave of feature updates for multimodal models. Multiple companies announced that their flagship multimodal models have added spatial understanding capabilities, enabling real-time inference of object positions, relative distances, and three-dimensional structures in images and videos. Unlike previous simple object recognition, the new features allow models to understand spatial relationships such as 'the cup on the table is about 15 cm to the left of the laptop' and provide actionable interaction commands. For example, a lightweight multimodal model launched by a domestic startup can analyze indoor scene layouts on a mobile phone with an error rate of less than 3%. In addition, another overseas AI company released a model variant supporting 3D point cloud input that can directly generate structured scene graphs from LiDAR data. These updates provide more precise perception capabilities for scenarios such as robot navigation, augmented reality, and smart homes, also laying the foundation for subsequent physical world interactions.
Spatial understanding improvements drive robot and AR interaction deployment
The launch of spatial understanding features has directly advanced applications in multiple verticals. In robotics, multiple service robot manufacturers have begun integrating multimodal models with spatial reasoning capabilities for automatic path planning, object grasping, and obstacle avoidance. For example, a warehousing and logistics company used the model's spatial inference capabilities to achieve zero-debugging shelf picking, reducing deployment time from two weeks to two days. In augmented reality (AR), the new model can recognize the user's surroundings in real time and overlay virtual information, such as identifying screw positions and marking removal sequences in maintenance scenarios. Some game companies have also started experimenting with this feature to create more natural virtual character interactions. However, running the new features imposes higher demands on inference resources, especially when deployed on edge devices, where model compression and quantization techniques still need optimization. Industry observers point out that spatial understanding is a key step toward general robot intelligence, but currently only leading vendors have mature productization capabilities.
Enterprise deployment cost optimization: from token wholesale to AI relay stations
As multimodal model features become increasingly rich, the cost pressure in enterprise deployment has become more prominent. To address this challenge, a new business model is rapidly maturing: offering token wholesale services through AI relay stations. Previously, enterprises directly interfacing with multiple model APIs often faced high unit prices and complex settlement. Now, third-party AI channel agent platforms centrally purchase inference quotas from multiple vendors and resell them to downstream customers at wholesale prices. According to feedback from multiple enterprises, the comprehensive cost of calling large model APIs through such token relay station models can be reduced by more than 40%, and enterprises can even obtain the cheapest token options on the market. At the same time, some large cloud vendors have also started building their own AI API relay stations, bundling their own models with other open-source models and providing unified scheduling and billing interfaces. In this model, enterprises do not need to worry about switching underlying models; they only need one API key to access multiple model options while enjoying bulk discounts. For example, an AI model aggregation platform (often called an AI hub) integrates dozens of mainstream and emerging models, allowing customers to switch on demand and automatically select the cheapest token path. This innovation both reduces enterprises' trial costs and promotes healthy competition among models. It is expected that by the end of 2026 to early 2027, token wholesale and AI relay station services will become standard for enterprise AI procurement, further pushing AI applications from experimentation to large-scale production.