Zhipu and MiniMax double down on AGI as SenseTime open‑sources vision model

This cluster of Chinese moves is one of the clearest signals yet that domestic frontier players are explicitly re‑orienting around AGI as the target, not just competitive LLMs. Zhipu’s “Touch High” plan commits the company to spend the next two years attacking long‑horizon tasks, autonomous agent systems, self‑training and mechanistic interpretability—precisely the areas most labs see as bottlenecks on the path to general intelligence. MiniMax’s founder going to zero salary “until AGI” and earmarking 5% of equity for staff and open‑source sends a strong internal and external signal that they view this as a decade‑scale mission, not a quick exit.

Meanwhile, SenseTime’s full open‑sourcing of SenseNova‑Vision, a unified model that rivals specialist systems across detection, segmentation, depth and multi‑view 3D, underlines how Chinese firms are using openness selectively to build ecosystem lock‑in. Combined with OpenRouter data showing Chinese models dominating global token usage, these steps suggest that China is not just catching up in LLM benchmarks but actively building alternative model families and tooling.

For the global AGI race, the implication is that we are firmly in a multipolar world. Ambitious, well‑funded Chinese labs are publicly staking out AGI roadmaps, betting on world‑model and agentic architectures, and increasingly open‑sourcing components that can be recombined by the broader ecosystem. That will accelerate innovation—but also complicate global coordination on safety and governance.