Holo3.1: Local computer use agents from 0.8B to 35B parameters

H Company, a French AI startup, has released Holo3.1, the latest iteration of its Vision-Language Models (VLMs) designed specifically for computer use agents. Building on the previous Holo3 release, the new family extends automation support from browser and desktop environments to mobile platforms, marking a significant expansion in capability. The models are fine-tuned from the Qwen 3.5 series and come in four sizes: Holo3.1-0.8B, Holo3.1-4B, Holo3.1-9B, and Holo3.1-35B-A3B (a 3B active parameters variant). Each variant is optimized for different deployment scenarios, with the largest model offering quantized checkpoints in BF16, FP8, NVFP4, and Q4 GGUF formats to enable local inference on consumer hardware. A key technical highlight is native function-calling support, allowing seamless integration with agent frameworks like LangChain or OpenAI-compatible APIs, which streamlines the development of autonomous agents that can control screens, click buttons, fill forms, and navigate apps across web, desktop, and mobile UIs.

Holo3.1's performance targets have been validated across benchmarks for computer use, UI grounding, mobile automation, and business workflows. By providing open-source weights under the Apache 2.0 license, H Company makes these models accessible to developers and enterprises that need private, on-device automation without sending data to external APIs. The inclusion of multiple quantization levels means users can trade off accuracy for speed and memory usage depending on their hardware—from edge devices (0.8B) to powerful workstations (35B). This flexibility could accelerate adoption in sectors like robotic process automation (RPA), accessibility tools, and software testing. Pre-trained checkpoints are available on Hugging Face, with quickstart guides on the company's hub. As AI agents become more prevalent, Holo3.1 offers a compelling, locally deployable alternative to proprietary cloud-based solutions.