Reportedly: China Merchants' LionRock AI Lab Open-Sources Lion-ER-35B-A3B Embodied Reasoning Model
Reported by Pandaily · not yet confirmed by the company or a second independent outlet. We update this page when it is.
China Merchants' LionRock AI Lab open-sourced Lion-ER-35B-A3B, a 3B-active embodied reasoning model built with a capability-driven data engine that beat its base on ten benchmarks. The source release was published on 2026-11-05, Pandaily reported.
China context
- Outside China
- Open weights · pandaily.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can access the model weights to experiment with embodied reasoning tasks, potentially reducing development time for robotics applications.
- For investors
- Investors should monitor whether this open-source release leads to commercial adoption or further investment in embodied AI startups.
China Merchants' LionRock AI Lab has open-sourced Lion-ER-35B-A3B, an embodied reasoning model with 3 billion active parameters. The model was developed using a capability-driven data engine and outperformed its base model on ten benchmarks, according to Pandaily.
The model uses a capability-driven data engine, which suggests a focus on targeted data curation for embodied reasoning tasks. The 3B active parameter count indicates a mixture-of-experts architecture, balancing performance and efficiency.
Developers outside China gain access to a new open-weights embodied reasoning model, potentially reducing the cost and time to build robotics applications. Competitors in the embodied AI space must now benchmark against Lion-ER-35B-A3B's reported performance.
Open-sourcing the model allows enterprises and researchers to integrate embodied reasoning capabilities without licensing fees, potentially accelerating development of robotics and autonomous systems.
Specific signals to watch include independent evaluations of Lion-ER-35B-A3B on embodied reasoning benchmarks and any downstream adoption in robotics projects. The model's performance on real-world robotic tasks remains to be verified.