The frontier is running out of data. China isn’t.
Real trajectories have no shortcuts.
Emerge sources, builds, and produces AI training data in China — licensed game trajectories, purpose-built RL environments, and custom agent, video, and speech data — delivered to frontier labs worldwide.
Licensed gameplay from China’s game industry.
We find the studios, clear the rights, integrate with the engine, and capture full player trajectories — video, inputs, state, tasks, and outcomes, frame-synchronized.
Explore Data ProductsBuilt to your model’s spec, not off a shelf.
RL environments and China-scenario collection — desktop and mobile agents, egocentric video, dialect speech — scoped around your model’s weakest task and proven in a paid pilot.
Work With UsEvery trajectory we ship is licensed at the source, schema-complete, and verified end to end.
Rights-cleared player trajectories from live and sunset Chinese games.
World ModelsResettable, parameterized environments with automatic reward and result verification.
RL ResearchDesktop and mobile agent traces across WPS, Feishu, DingTalk, and e-commerce back ends.
Computer-Use AgentsFirst-person hand-operation video: packing, sorting, assembly, and tool use.
Embodied AI & VLADialect and scenario speech — segmented by region, age, environment, and device.
Speech & AudioVideo, inputs, engine state, task goals, and outcomes — frame-synchronized.
Game AgentsSuccess, failure, and recovery trajectories from real human operators.
Offline & Online RLThree ways we supply the data frontier AI needs
A game data supply network across China's studios.
We source mid-size studios, indie teams, and sunset titles, run rights diligence, license gameplay for AI training, and instrument the engine to capture full player trajectories.
Explore Game DataSimulation & RL environments built to be trained on.
Resettable, parameterized, fully instrumented environments with clear task goals, automatic rewards, and complete engine-state output — sold as trajectories, task generators, or executable environments.
Explore EnvironmentsCustom collection for Chinese scenarios.
Agent traces, egocentric video, dialect speech, document understanding, and e-commerce data — produced to your spec through a paid pilot, not stockpiled and resold.
Start a PilotFrom China's studios to your training runs.
How we work.
Your model's weakest task.
Our next pilot.
Tell us what your model can't do yet. We'll scope a paid pilot: 20–50 tasks, 50–200 trajectories, a data schema, and a quality & licensing report.
Start a Pilot