Holo4: Powering generalist computer-use agents

Holo4 has been announced as a new series of generalist agentic models capable of seamlessly interacting with various software interfaces across desktop, web, and mobile platforms. Available in 27B and 35B-A3B sizes, Holo4 utilizes both supervised and reinforcement learning trained on diverse environments, making it adept at executing complex business workflows. With a notable performance score of 85.2% on the OSWorld benchmark and a cost-efficient task pricing at $0.08, Holo4 demonstrates superior capabilities over existing models such as Qwen 3.8, especially in environments requiring multiple interface types.

The significance of Holo4 lies in its interoperability; it can execute tasks by clicking graphical user interfaces, writing and running code, and utilizing APIs—all without needing to switch models for different platforms. Unlike conventional models that are often tailored to a single interface, Holo4 integrates multiple functionalities, greatly enhancing its utility in real-world applications. Its training involved a remarkable Agentic Task Factory, generating around 10,000 tasks that facilitate cross-platform operability. This model marks a step forward in the AI/ML community, indicating a shift towards more versatile and cost-effective solutions for automating complex tasks across diverse computing environments.