Holo4: A New Generation of Generalist Computer‑Use Agents

Holo4 is released in two variants: a 27‑billion‑parameter dense model and a 35‑billion‑parameter mixture‑of‑experts version (35B‑A3B). Both are accessible through the H Models API and share the same weights, meaning a single deployment can handle desktop GUIs, web interfaces, Android environments, code sandboxes and enterprise APIs.

Interface‑agnostic operation

Unlike many agentic systems that specialize in one interaction mode, Holo4 decides at runtime whether to click and type on a screen, write and execute its own script, or call an MCP or API tool. This flexibility mirrors real‑world business workflows where a task often requires a mix of GUI actions, custom code and service calls.

On the OSWorld 2.0 benchmark, Holo4‑27B achieves 61.7 % success, close to the leading closed model Opus 5.5 at 81.8 %, while the 35B‑A3B variant reaches 30.9 %. Importantly, these results are obtained with orders of magnitude fewer parameters and a substantially lower cost per run, as measured by token‑based pricing on the H Models API.

Why it matters for GPU / AI infrastructure: The reduced parameter count translates to lower memory footprint and inference latency, allowing Holo4 to run efficiently on mid‑range GPUs or even CPU‑only instances. This cost‑effective profile makes large‑scale agentic automation feasible for enterprises that need to orchestrate complex workflows without investing in massive AI clusters.