Holo3.1、ローカルAIエージェント向け量子化に対応、Androidで79.3%を達成

holo3.1 expands computer-use agent capabilities to mobile environments and introduces the first quantized checkpoints (fp8, q4 gguf, nvfp4) for local inference. the 35b-a3b model improved performance on androidworld from 67% to 79.3%, while smaller 0.8b, 4b, and 9b models were released for cost-effective deployment. local execution on consumer hardware and dgx spark delivers up to 2x faster step times and full privacy. this signals a shift towards robust, private computer use agents running on-device rather than in the cloud.