The Chinese AI evaluation group SuperCLUE has published the September edition of its EmbodiedCLUE-VLA ranking, which assesses the ability of AI models to act as the cognitive core of a robot, planning and reasoning while performing physical tasks.
Among Chinese developments, Alibaba's Qwen3.8-Max-0902 and ZTE's Nebula-EmbodiedBrain secured the first place, achieving identical overall scores of 77.48.
The testing covers four key areas: basic perception, which is divided into temporal, object, and spatial perception; visual reasoning, including mathematical, temporal, spatial, and logical reasoning; interaction and planning, consisting of task planning and trajectory planning; and embodied safety, which checks adherence to ethical norms, physical safety, and privacy protection.
SuperCLUE considers models whose scores differ by one point as a tie and includes international models only for reference outside the main ranking.
How the Leaders Achieved the Ranking
The two leaders achieved the same overall result through different paths. The Qwen3.8-Max-0902 model showed the best results among Chinese models in basic perception (83.33) and visual reasoning (84.72). Meanwhile, Nebula-EmbodiedBrain scored 97.92 in embodied safety, surpassing the Alibaba model, which scored 89.58.
Both models received an identical score of 39.29 in the interaction and planning category, but the ZTE model demonstrated greater strength in task planning (64.29 versus 28.57 for the Alibaba model), while the Alibaba model was stronger in trajectory planning (50 versus 14.29).
According to the report from the Chinese tech publication MyDrivers, the Alibaba model features the ability for spatio-temporal memory, allowing it to remember unfinished work after a robot task interruption. The ZTE version, conversely, is oriented towards on-device deployment and adaptation to the robot's hardware, making it more suitable for operation on physical machines. ZTE previously released EmbodiedBrain 1.0—a vision and language model for embodied task planning, with 7B and 32B weights published on Hugging Face.
Following in the ranking are open-source models with 10 billion parameters or less. Xiaomi's MiMo-Embodied-7B took second place among Chinese models with a score of 56.95 and topped a separate SuperCLUE ranking for models under 10B. It was followed by Alibaba's RynnBrain1.1-9B with a score of 50.33, then BAAI's RoboBrain2.5-8B-NV with 46.36, and Tencent's HY-Embodied-0.5, a 4B model that scored 29.14. Both leaders in this group are closed models available via API.
Planning remains the weakest area among Chinese developers. No Chinese model scored above 40 in the interaction and planning category, and four open entries scored 14.29 or lower. Among comparison models, GPT-6 Astra scored 86.75 overall, and Gemini-3.8-Flash scored 82.12, while Gemini-Robotics-ER-2-Preview, focused on robotics, received 70.86.
MyDrivers noted that as more companies create embodied brains, rankings based on different test sets often yield contradictory results, and the capabilities of these models will largely determine the scope of practical work future robots can perform. SuperCLUE previously published editions of this ranking in January and February 2026.
