The LionRock AI Lab, which operates under the China Merchants Group Advanced Technology Research Institute, has released Lion-ER-35B-A3B. According to a report by QbitAI on November 4, the model is an embodied reasoning vision-language model designed to assist robots in understanding three-dimensional space, object states, and task progress. The lab has made the weights available on Hugging Face under the Apache 2.0 license.
Lion-ER, which stands for LionRock Embodied Reasoner, is a mixture-of-experts model. It features approximately 35 billion total parameters, with about 3 billion active per token. The lab's model card states that the model was fine-tuned from Qwen3.6-35B-A3B.
The lab reported that the model improved upon its base model across all ten public embodied and visual reasoning benchmarks it tested. It reached an equal-weight average of 65.29, with scores averaged over four sampled answers per question. Specific results reported by the lab include a Where2Place spatial placement score of 64.28, which is a 9.30-point increase over the base model, and a score of 64.38 on the ERQA embodied reasoning benchmark, an increase of 6.19 points. Additionally, it scored 89.33 on the SAT spatial reasoning benchmark, while the trajectory understanding subcategory of ERQA rose 12.12 points to 68.18.
The lab argues that simply collecting more robot data does not automatically result in increased capability. Instead, the team developed what it calls a capability-driven embodied data engine. The team mapped embodied reasoning into 39 finer skills across eight top-level domains, including entities, space, time, physics, affordances, tasks, social context, and evidence. Each skill was converted into concrete training questions that specified the scenes, observations, questions, and evidence required for a verifiable answer. Questions were dropped if the visual evidence was insufficient.
Human reviewers checked hard samples that the models repeatedly failed to answer correctly. The resulting supervised training set contains 312,380 examples, which were weighted by capability needs rather than the volume of the source. The team used an evaluation-driven loop to feed failures back into data production. After ruling out labeling or scoring errors, the team drew new material with the same reasoning pattern from a pool of about 4.3 million records across 31 data sources. For instance, when the model confused being higher in an image with being higher in 3D space, the team added scenes with multi-level surfaces.
The model card notes that Lion-ER is intended for research purposes. It has not been validated for autonomous robot control or safety-critical decisions. Furthermore, the card states that benchmark results may not transfer to new environments. The lab has also applied this same recipe to smaller 4B and 9B backbones.
