TL;DR: Holo3.1 is a new series of computer-using AI models with enhanced robustness across web/desktop/mobile, quantized checkpoints for local inference, and multiple sizes for cost/performance balance.
Summary: The Holo3.1 series improves on Holo3 with three key enhancements: robustness across diverse environments, native function calling for agent frameworks, and quantized checkpoints (FP8, Q4 GGUF, NVFP4) for local on-device inference. Four sizes (0.8B, 4B, 9B, 35B-A3B) are offered for deployment from ultra-lightweight to top-tier. Benchmark improvements are shown in mobile automation (AndroidWorld).
Why it matters: This release advances universal, private, locally executable computer-using agents. AI builders should evaluate the quantized checkpoints for deploying local agents on edge devices.
Source: bestblogs.dev
原文 (Original):
📌 One-Sentence Summary Holo3.1 是一系列新的计算机使用 AI 模型，提升了在网页、桌面和移动环境中的鲁棒性，引入了用于本地推理的量化检查点，并提供了多种尺寸以权衡成本与性能。 📝 Summary 本文宣布发布 Holo3.1 系列计算机使用智能体，该系列基于之前的 Holo3 模型构建。关键改进体现在三个方面：跨多种环境（网页、桌面和移动）的鲁棒性增强；通过原生函数调