Liquid AI released local CPU encoders, Ai2 accelerated continent-scale satellite inference, and BYD is preparing to show its first humanoid robot prototype publicly in August.
Liquid AI released two open-weight encoders with 230 million and 350 million parameters. They support an 8,192-token context and are designed for classification, routing, extraction and scoring. In the company's own measurements, the 230M model processed an 8,192-token input in about 28 seconds, roughly 3.7 times faster than ModernBERT-base. The comparison averages five seeds across 17 tasks, but the post does not identify the measurement hardware.[1]
Ai2's OlmoEarth Platform combines CPU-based data acquisition, GPU inference and CPU-based postprocessing in one pipeline. In its North America wildfire-risk example, roughly 19,600 CPUs and 994 GPUs ran in parallel, reducing an estimated 4,737-hour process to 30.5 hours for a 155-fold speedup. The job processed dozens of terabytes of imagery. Because the post omits the hardware, hourly prices, licensing terms and general availability date, the result remains an infrastructure example.[2]
BYD confirmed that it will publicly show its first humanoid robot in early August through its Di Space retail network. It has disclosed no specifications, price, availability or partners; the event will be the first showing of a physical prototype rather than a concept announcement. Liquid AI and Ai2 describe operating environments and measurements today, while those details do not yet exist for BYD's robot. Together, the developments place AI workloads under different conditions on local CPUs, large scientific infrastructure and physical product floors, rather than one chat surface.[1], [2], [3]
