Primotion's Paimon-0 cuts action-chunk inference from about 3,000 ms to 300 ms, company says

Primotion, a Shenzhen startup founded in July 2026 to build chips for embodied AI, says its new embodied foundation model Paimon-0 cuts inference for a single action chunk from about 3,000 ms to 300 ms. The company made its first public appearance with the model rather than with silicon, according to a QbitAI report published on Tencent News and an earlier report by The Beijing News.

Benchmark results (self-reported)

Primotion claims Paimon-0 ranks first on several public robot benchmarks. All figures below come from the company and have not been independently verified.

Benchmark	Result
RoboColiseum leaderboard	0.7754 (first overall; also first in general manipulation and disturbance-adaptation sub-rankings)
RoboCasa365	58.1 overall
VLA-Arena	0.73 average success rate
Dual-arm robot policy challenge	46.599 points (won)

Architecture

Paimon-0 is built around a world action model. The model predicts future states, and an inverse dynamics model turns those states into actions. Rendering full future video is too slow for real-time robots, so Paimon-0 uses query tokens to distill only core features, such as object 3D pose, semantics and affordance, before predicting actions on the device.

Primotion says this reduces single action-chunk inference from about 3,000 ms to 300 ms. The company stresses that this figure is chunk inference time, not the period of the low-level control loop.

Pretraining relies heavily on first-person human manipulation video supplied by partner EgoScale. Parameter count, context window and pricing are not stated.

Why a chip company started with a model

CEO Ren Qingyuan, a former vice president at a RISC-V processor company, argues that a team must first run the full loop of data, model, agent and evaluation on real customer tasks, then use those workloads to define the chip. Primotion splits work between a cloud planner it calls the meta brain and an on-device local brain that handles high-frequency feedback from vision, touch and force.

Hardware progress

Hardware work has already begun. By adding an FPGA alongside a robot's main controller to offload pre- and post-processing, Primotion says it cut worst-case timing error on a 1 kHz EtherCAT control cycle from 30 microseconds in software mode to 152 nanoseconds. Much of that logic is meant to move into the company's own system-on-chip, which it plans to start delivering in 2028.

Roadmap

• Primotion says it will not build robot bodies.
• It plans to open-source its progress model, Paimon-Eval, in November or December.
• It expects to release a second foundation model trained on first-person data in about three months.

What this means

The 300 ms figure describes how long one action chunk takes to infer, and Primotion explicitly separates it from the control loop period. The two numbers measure different things and should not be compared directly. The benchmark rankings are company-reported, so the planned open-sourcing of Paimon-Eval is the first step that could let outsiders check how progress is measured. Primotion's stated strategy is to use workloads from a deployed model to define its SoC, and the FPGA timing result is an early hardware data point for that plan. Delivery of its own chip is not planned until 2028.