H Company Launches Holo4 Generalist Computer-Use Agent Models

French AI company H Company has released Holo4, a new family of generalist agentic models built to operate across every software interface at once: graphical UIs, code, MCP servers, and traditional APIs. The release is the latest sign that computer-use agents are moving past single-interface specialization toward models that handle real, multi-step business workflows end to end.

One model for every interface

Holo4 ships in two sizes: Holo4-27B (dense) and Holo4-35B-A3B (Mixture of Experts), alongside Holotron4 Nano, an updated version of the company's earlier Holotron 3. All three are available through the H Models API, with weights published on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats.

The defining feature is interface-agnostic design. Most agentic models are trained for a single interface: GUI-focused models are blind without a screen, while tool-calling models stall in front of an app with no API. Holo4 runs the same model on desktops, the web, Android, in code sandboxes, and against business APIs, choosing clicks, code, or MCP tool calls as the task demands.

Competitive at a fraction of the cost

Built on a Qwen base, Holo4 is competitive with frontier systems while using far fewer resources. On OSWorld 2.0, the desktop-control benchmark, Holo4-27B scores 61.7% against 81.8% for Anthropic's Opus 5.5, but with orders of magnitude fewer parameters and much lower compute cost. The company open-sources every trajectory behind its public benchmark scores, so teams can replay and debug each step.

The model was trained with supervised and reinforcement learning on environments generated by the company's Agentic Task Factory, an internal pipeline that builds interactive, verifiable tasks from documentation alone. So far it has produced roughly 10,000 tasks across web apps, MCP servers, and desktop environments.

Why it matters for builders

For AI builders and automation engineers, Holo4 collapses the interface-fragmentation problem that makes agent deployment expensive and brittle. Instead of routing tasks to different models per interface, teams can run one model that handles screens, code, and MCP tools uniformly, and inspect its open-sourced trajectories to understand agent behavior.

The NVIDIA Nemotron Coalition partnership points to where open-weight agent models are heading: Holotron4 Nano is built on Nemotron 3 Nano Omni, applying the same post-training recipe to a smaller, faster base. As computer-use agents move from demos into production, interface-agnostic models like Holo4 lower the barrier for teams automating real workflows.