Primotion’s Paimon-0 model tops robot benchmarks and cuts action-chunk inference to 300 ms

Primotion, a Shenzhen startup founded in July 2026, announced its embodied foundation model Paimon-0, which the company says ranks first on several public robot benchmarks. The scores—0.7754 on the RoboColiseum leaderboard, 58.1 overall on RoboCasa365, a 0.73 average success rate on VLA-Arena, and 46.599 points in a dual-arm robot policy challenge—are self-reported by Primotion. source

The model uses a world-action architecture: a world action model predicts future states, which an inverse dynamics model turns into actions. To avoid the latency of rendering full future video, Paimon-0 employs query tokens that distill core visual features such as 3D object pose, semantics and affordance before action prediction. Primotion says this reduces inference for a single action chunk from roughly 3,000 ms to 300 ms.

Primotion adds an FPGA alongside a robot’s main controller to offload pre- and post-processing. The company reports that worst-case timing error on a 1 kHz EtherCAT control cycle fell from 30 µs in software mode to 152 ns with the FPGA. Much of this logic is intended to migrate onto a dedicated system-on-chip, which Primotion plans to begin delivering in 2028.

CEO Ren Qingyuan, a former vice-president at a RISC-V processor firm, argues that running the full loop of data, model, agent and evaluation on real tasks first defines the chip workload. Primotion will not build robot bodies. It plans to open-source its evaluation model, Paimon-Eval, in November or December and expects to release a second foundation model trained on first-person data in about three months.