九识建成首个L4万卡集群，无人驾驶进入多模态大模型新范式
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Jiushi says it has built the L4 autonomous-driving industry’s first 10,000-plus-accelerator cluster, totaling nearly 15,000 cards, to scale its APEX multimodal foundation model toward 100 billion parameters. The system combines fleet data, cloud models, onboard VLA models, and a safety-agent fallback to improve long-tail urban driving.
- APEX trains on data from more than 30,000 vehicles operating across 300-plus cities, including 270 million kilometers of real-world L4 driving.
- Its multimodal training blends driving data with language, video, traffic knowledge, and human decisions to support prediction, planning, simulation, and cross-city generalization.
- Jiushi’s architecture assigns routine driving to onboard models, uses cloud VLA assistance for harder low-speed conditions, and invokes a safety agent for extreme 0–5 km/h scenarios.
- Improved foundation models are distilled back into lighter vehicle models, creating a feedback loop between fleet operations, cloud training, evaluation, and deployment.
Sources (1)
九识建成首个L4万卡集群，无人驾驶进入多模态大模型新范式
TL;DR - Jiushi says it has built the L4 autonomous-driving industry’s first 10,000-plus-accelerator cluster, totaling nearly 15,000 cards, to scale its APEX multimodal foundation model toward 100 billion parameters. The system combines fleet data, cloud models, onboard VLA models, and a safety-agent fallback to improve long-tail urban driving.
- APEX trains on data from more than 30,000 vehicles operating across 300-plus cities, including 270 million kilometers of real-world L4 driving.
- Its multimodal training blends driving data with language, video, traffic knowledge, and human decisions to support prediction, planning, simulation, and cross-city generalization.
- Jiushi’s architecture assigns routine driving to onboard models, uses cloud VLA assistance for harder low-speed conditions, and invokes a safety agent for extreme 0–5 km/h scenarios.
- Improved foundation models are distilled back into lighter vehicle models, creating a feedback loop between fleet operations, cloud training, evaluation, and deployment.