LionRock AI Lab open-sources 35B embodied reasoning model for robotics

China Merchants' LionRock AI Lab has released the Lion-ER-35B-A3B model, an embodied reasoning vision-language system aimed at helping robots interpret 3D space, object states and task progress, according to a report on Pandaily. The weights are hosted on Hugging Face under an Apache 2.0 license.

Lion-ER (LionRock Embodied Reasoner) is a mixture-of-experts architecture with roughly 35 billion total parameters and about 3 billion active per token, fine-tuned from Qwen3.6-35B-A3B. The lab says it improves on its base across ten public embodied and visual reasoning benchmarks, achieving an equal-weight average score of 65.29. Notable results include a Where2Place spatial placement score of 64.28 (up 9.30 points) and an ERQA score of 64.38 (up 6.19 points).

The team built a capability-driven embodied data engine rather than scaling raw data. They mapped reasoning into eight top-level domains and 39 finer skills, generating 312,380 supervised examples weighted by capability need. An evaluation-driven loop then feeds failures back into data production from a pool of about 4.3 million records across 31 data sources.

The model card notes that Lion-ER is intended for research, has not been validated for safety-critical decisions or autonomous robot control, and that benchmark gains may not transfer to new environments. The same methodology has been applied to smaller 4 B and 9 B backbones.

Our read: The open-source release lowers the barrier for Chinese developers to experiment with high-capacity embodied reasoning, but the modest benchmark gains and explicit safety caveats suggest it is a research-grade tool rather than a production-ready robot controller.