Unverified50% confidenceBenchmarkExact time
The training scale is 4096 games in parallel, about 330,000 agent steps per second end-to-end, completing about 19 billion agent steps in about 30 hours on a single Lambda GH200.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/9/2026
First Seen
Valid until: 2/7/2027
Sources
Related Entities
Related Claims
UnverifiedAgent training is moving from completing a task in a few dozen steps to running for several hours and processing nearly a million tokens.77% similarUnverifiedThe simulator has a C++ version that can complete 10 million steps of training in about 6 minutes, roughly 28,000 steps per second.75% similarUnverifiedSingle-CPU training took about 20 minutes, and the agent has about 475,000 parameters.73% similarUnverifiedTraining used a compiled Numba simulator with throughput several orders of magnitude higher than real replays; the agent makes a decision every 4 server ticks, about 133 milliseconds.72% similarUnverifiedThinker-WM's training corpus totals approximately 20K hours of first-person video, 10K hours of real robot data, and 1K hours of simulation data, of which about 4K hours of real robot data has aligned depth supervision.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1152063API
curl https://kongchang.com/api/v1/knowledge/claims/1152063MCP
get_claim(id=1152063)