Inverse Matrix, BAAI, vLLM and Peking University Release Physical AI Inference System PhysisServe

On October 21, 2026, Inverse Matrix, together with the Beijing Academy of Artificial Intelligence (BAAI), vLLM, and Peking University, released PhysisServe, a physical AI inference system described as the first to combine intelligence with high-throughput output. It is the first system to enable a 10-billion-parameter world model to generate at a real-time frequency of one kilohertz, one to two orders of magnitude faster than existing frontier approaches.

Physical AI inference has faced a fundamental trade-off: larger models describe the world more accurately but require more computation per step, while autoregressive denoising relies on serial dependencies that cannot be parallelized, making generation slower. The release suggests that, given sufficient compute, physical intelligence foundation models could scale to tens of billions, hundreds of billions, or even trillions of parameters while still producing real-time outputs at task-required frequencies.

Physical AI systems can already perform actions in controlled environments, but moving from stable laboratories to real-world settings full of noise, interference, and constant change remains a long way off. Inverse Matrix focuses on building general world foundation models for the real physical world, aiming to develop general physical intelligence that can understand, predict, and act across scenarios. The company says it hopes to create a physical intelligence foundation that can accumulate and reuse experience across scenarios, enabling systems to understand, predict, and influence the real world at the speed and accuracy required by actual tasks.