Physix Frontier · News Briefing Card (QbitAI · Jul 23, 2026)

Tars AI Launches AWE 3.5 Embodied Native Model for Industrial Multi-Task Generalization

KEY FACTS

Tars AI unveiled its embodied native foundation model AWE 3.5 at WAIC.
The model was trained on over one million hours of human-centric data.
It adopts a One Model architecture that unifies pre-training of vision, language, and action.
New tasks require only hour-level data collection to reach a usable state.
On-site demonstrations showed multi-scenario robot collaborative operations, including automotive wiring harness assembly.

KEY DATA

Over 1 million hoursTraining Data Scale

Hour-levelNew Task Collection Time

PHYSIX OBSERVATION

AWE 3.5 marks a critical turning point for embodied intelligence moving from single-point demos to a general-purpose foundation. Its core breakthrough lies in replacing VLA stitching with a native multimodal architecture, enabling 'chain-of-thought' reasoning in the physical world. When the cost of adapting to new tasks drops to the hour level, the industry barrier will shift from algorithms to high-quality data closed loops and coordinated dexterous hardware. This signals that the embodied scaling dividend period has arrived, and 2027 may become the watershed for large-scale factory deployment.

Source: QbitAI report