AI advancements: Google's Gemini 4 Pro and Unitree's robotics model reach new benchmarks

Advancements in artificial intelligence are accelerating across both large language models (LLMs) and embodied intelligence. A model appearing in the Arena under the pseudonym “gemini-3.8-flash” is widely believed to be Google’s unreleased Gemini 4 Pro. Preliminary testing suggests the model achieves state-of-the-art (SOTA) performance across several benchmarks, including DeepSWE v1.1 for coding and OSWorld-2.0 for computer operations, outperforming competitors like GPT-6 Astra in specific tasks.

In the field of robotics, Chinese company Unitree has introduced UnifoLM-WLA-1.0, a 6B-parameter general-purpose foundation model for humanoid robots. The model has achieved SOTA results in seven open-source benchmarks and demonstrates strong spatial understanding, with some scores exceeding those of GPT-6 Astra in benchmarks like Where2Place and EmbSpatial. Unitree’s model is designed to coordinate 64 different tasks, ranging from household chores like folding towels to complex full-body movements, utilizing an interaction-centric world modeling approach to predict environmental changes during physical tasks.