AnswersPublished18 sources
Why Primotion Launched a Robot Model Before Its Chips
Primotion introduced Paimon 0 before launching finished silicon, using model and runtime development to inform its chip plans.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Primotion introduced Paimon 0 before launching finished silicon, using model and runtime development to inform its chip plans.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How did Shenzhen embodied-AI chip startup Primotion, founded in July 2026, make its public debut with Paimon-0 rather than silicon, and what. Article summary: Primotion debuted as a chip startup with a model, Paimon-0, because its stated strategy is to develop algorithms, runtime software and hardware together rather than design silicon in isolation.[1][6] There is an importan. Topic tags: general. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clickbait thumbna
Primotion made its public debut with Paimon-0, an embodied-AI model, while its proprietary chips were still in early development. The model-first approach fits the company’s stated plan to build algorithms, runtime software and computing hardware together. But the benchmark story needs a caveat: media reports describe Paimon-0 as a RoboCasa365 leader, while the benchmark’s October 19 leaderboard placed it fourth.1
3
8
Primotion’s first public result was a model, not a chip
Primotion was reported as founded in July 2026. Its first public business update came on October 16—not October 26—and introduced Paimon-0 alongside nearly RMB 100 million in purchase intentions. Those were intentions, not completed sales. The company’s own chips were still at an early development stage.3
4
That sequence reflects Primotion’s stated product strategy: treat the model and runtime as part of the chip solution, rather than define hardware separately from the work robots need to do. The company says it is developing models, runtime software, computing platforms and specialized chips in coordination.1
3
What the Paimon-0 benchmark reports say
Coverage of Primotion’s claims reports a 0.7754 overall score on RoboColiseum, a 0.6613 score in general manipulation, and a lead in the perturbation-adaptation sub-ranking. The same coverage reports 0.73 average success on VLA-Arena and 46.599 points in a dual-arm robot policy challenge.2
7 These are reported results, not independent evidence that the model will perform equally well on physical robots in everyday settings.
RoboCasa365 is the clearest qualification. Media reports describe Paimon-0’s 58.1 overall score as a first-place result. But the RoboCasa leaderboard, updated October 19, lists Paimon-0 at No. 4 with that same score. Its entry also records 79.3% atomic-seen, 55.8% composite-seen and 36.7% composite-unseen success.2
8
19 The first-place claim should therefore not be repeated without noting the leaderboard snapshot’s different ranking.
Why lead with a model?
Paimon-0 is described as a hierarchical asynchronous world-action model (WAM). It brings vision, language, robot state and actions into a shared representation, with computation coordinated across different timescales. The stated aim is to reduce how much action execution must wait for complex reasoning.3
That is relevant to Primotion’s hardware plans: the company says execution evaluation and cloud–device cooperation can help reveal robots’ real computing requirements and inform chip design.18 The available sources support this broad co-design rationale; they do not independently establish every technical detail in the more elaborate architecture claims.
Keep the latency claim in its lane
An October 26 report says Primotion’s action-chunk inference reaches 300 milliseconds.7 That figure concerns model inference for an action chunk; the available source excerpt does not establish a verified 3,000-millisecond baseline or a change in low-level control-loop timing. Those are different measurements, so the reported latency should not be presented as proof that the robot’s control loop became ten times faster.
The evidence supports a clear takeaway: Primotion is using a model-first debut to demonstrate its approach before its chips are ready. Its reported scores provide an early signal, but the RoboCasa ranking discrepancy and limited detail on latency measurements make careful attribution essential.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Why Primotion Launched a Robot Model Before Its Chips"?
Primotion introduced Paimon 0 before launching finished silicon, using model and runtime development to inform its chip plans.