Researchers at the Merchant Bureau's Lion Mountain Artificial Intelligence Lab have introduced "Yuwu™ Lion-ER-35B-A3B," a new embodied reasoning vision-language model. This model utilizes a Mixture-of-Experts architecture, activating approximately 3 billion parameters per token out of a total of 35 billion. Instead of relying on massive datasets, the model employs a capability-driven data engine that focuses on designing training data around specific skills the model needs to acquire. This approach has led to improvements across ten mainstream embodied AI benchmarks, with some metrics surpassing larger models. AI
IMPACT This model's novel data-centric approach could lead to more capable and efficient robots in real-world physical tasks.
RANK_REASON New embodied reasoning model release from a research lab. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
