Our research moves along three axes: bringing models off the cloud, letting data generate itself, and turning general intelligence into domain expertise. Together, they shape our answer to what comes after the foundation-model era.
Directions
端侧大模型部署应用
Compressing and quantizing billion-parameter language models for deployment on edge devices. Native offline inference, no cloud dependency — bringing large models to environments where privacy is non-negotiable and connectivity is unreliable.
数据仿真
Synthesizing high-fidelity training data through procedural generation and physics-based simulation. Breaking the cost ceiling of real-world collection — covering edge cases, rare events, and extreme conditions at unbounded scale.
专业领域大模型
Building expert-grade models for verticals through knowledge injection and fine-tuning on top of foundation backbones. More precise than general-purpose models, more adaptive than rule-based systems — trusted intelligence for finance, healthcare, and industrial applications.