Qianzhang Intelligence's Q0-Preview has claimed the top spot on the RoboDojo Sim dual-arm robot simulation evaluation leaderboard, surpassing Anthropic's Fable 5.2 with a task success rate of 68.87% and an overall score of 74.2—making it the first solution on the leaderboard to exceed an overall score of 70. Rather than adopting the VLA or WAM technical paradigms, this solution takes an Agent-based approach, building on Anthropic's previous-generation model Opus 5.5 as its foundation. Its accuracy rate far exceeds the 39.4% level that Anthropic publicly reported for that model. The core idea is to reframe embodied manipulation as a spatial computation problem—determining the contact position and pose at each manipulation stage. Leveraging the spatial perception and programming capabilities of multimodal large models, the system maps contact targets into the world coordinate system through pixel localization combined with camera parameter calculation. Three spatial reference points are then used to uniformly represent the gripper's target pose and contact information, which downstream motion planning components convert into continuous actions—forming a closed loop of observation, computation, execution, and feedback. An accompanying Harness system governs the interaction between the model and each module across four dimensions: clarifying execution boundaries and tool invocation, accumulating general manipulation experience through simulation trial and error, supplementing task requirements to clarify completion conditions, and managing context and task memory—thereby ensuring smooth system operation. This solution performs exceptionally well in areas such as dual-arm synchronous and coordinated manipulation, precise manipulation, adaptive adjustment, and memory, with some tasks even surpassing expert teleoperation. The solution's code is currently open-sourced, and the team will subsequently release a technical report and optimized code, along with additional real-machine experiments. At present, embodied manipulation based on foundation models still faces challenges: inference efficiency and cost remain high, the mechanisms underlying capability acquisition are yet to be clarified, and long inference latency makes it difficult to adapt to dynamic real-world environments. A new path—one that combines general large models with embodied domain knowledge constraints—is driving embodied intelligence toward a new development inflection point.