Shenzhen robot-chip startup Primotion unveils embodied model Paimon-0, says action-chunk inference cut to 300ms, a tenth of before (report)

Verified (fact)

Primotion, a three-month-old embodied AI chip startup in Shenzhen, China, has unveiled Paimon-0, an embodied foundation model it trained itself. The company chose to train its own model before designing robot chips. According to the report, it cut the time to infer an action chunk (the batch of movements a robot outputs at once) to 300ms, about one-tenth of previous levels. It also laid out a roadmap leading to an FPGA-based control chip. How the model will be released and on what terms has not been confirmed.

Needs fact-check (claim)

The company says Paimon-0 ranked first on several embodied AI leaderboards. QbitAI reported that the model outperformed GPT-6 Astra, but the excerpt does not say on which evaluation.

Why it matters (interpretation)

This is a case of a robot-chip company building its own model first to set the direction of its hardware design. If the 300ms inference time holds on real robots, competition could shift toward reducing the latency that weakens VLA models in real-time control.

Sources

• Press Shenzhen Embodied-Chip Startup Primotion Unveils Paimon-0, Cutting Action-Chunk Inference to 300 ms Pandaily
• Press 成立3个月便反超GPT-6 Astra！深圳黑马首次亮相 量子位