French AI developer H has released Holo 4 models designed to let AI agents work across command lines, APIs and graphical interfaces, including clicking, scrolling and typing. The models aim to address a longstanding weakness in AI: interacting with desktop software built for people, which could make agents more useful for tasks that require operating applications directly.

Holo 4 is based on Alibaba’s Qwen models and was trained with supervised learning and reinforcement learning; H has also updated its Holotron model, based on Nvidia’s Nemotron 3. The company says Holo 4 beats some larger models on computer-use benchmarks, though its tasks can cost more, partly because it uses more reasoning tokens. Its open weights can run at 4-bit precision on a 24 GB RTX 3090, and H has released a compatible GGUF version and its own agent harness; other developers are also pursuing computer-use models.