Not relying on piling up data, enabling robots to understand the real physical world | China Merchants Group Lion Rock Lab releases "YuWu™ Lion-ER-35B-A3B" embodied reasoning model

What happened: China Merchants Group's Lishan AI Lab released the embodied reasoning vision language model Yuwu Lion-ER-35B-A3B, built on a MoE architecture with about 35B total parameters and about 3B activated per token. Key numbers: 312,380 supervised training samples; 10 public benchmarks, all Why it matters: The lab says the model follows a capability-driven data engine rather than scaling training data, and reports several core abilities surpassing larger models. Weights are open-sourced on Hugging Face in BF16 Safetensors format, compatible with Transformers, vLLM and SGLang.

China Merchants Group's Lishan AI Lab released the embodied reasoning vision language model Yuwu Lion-ER-35B-A3B, built on a MoE architecture with about 35B total parameters and about 3B activated per token.

The lab says the model follows a capability-driven data engine rather than scaling training data, and reports several core abilities surpassing larger models. Weights are open-sourced on Hugging Face in BF16 Safetensors format, compatible with Transformers, vLLM and SGLang.