Hcompany’s Holo4 Targets Computer Use Across GUIs, Code, MCP and APIs

Hcompany describes Holo4 as a new series of agentic models offered in two sizes: a 27B dense model and a 35B-A3B Mixture of Experts model. [1] Both models are available through the H Models API, while the release also includes an updated Holotron 3 called Holotron4 Nano. [1] Holo4 can click and type on screens, write and execute its own code, and invoke MCP or API tools, choosing an interface according to the task. [1] The same model runs on desktops, the web, Android, in a code sandbox and against business APIs, without requiring users to select a separate model for each platform. [1]

Hcompany says Holo4 was trained with supervised and reinforcement learning across a large set of environments and tasks, including tasks generated by its Agentic Task Factory. [1] The company says real business tasks can require multiple approaches, while GUI-focused models depend on a screen and tool-calling models can be blocked when an application lacks an API. [1] On OSWorld 2.0, Holo4 27B scored 61.7%, compared with 81.8% for Opus 5.5, while Holo4 35B-A3B reached 30.9%. [1] Hcompany says the Holo4 models improve significantly over their Qwen base and operate with orders of magnitude fewer parameters and at a much lower cost than the strongest closed models. [1]

Holo4’s main proposition is interface consolidation: one model is intended to move between screen interaction, code execution and tool calls rather than forcing a platform-specific choice. The OSWorld 2.0 result shows a mixed competitive picture: Holo4 27B’s 61.7% trails Opus 5.5’s 81.8%, but Hcompany pairs that gap with claims of fewer parameters and lower cost. For teams, the key evaluation is whether cross-interface flexibility offsets the benchmark deficit in their actual workflows. What the sources don't address: how Holo4 performs on reliability, security and total task cost in independently measured production deployments.