Holo4: powering generalist computer-use agents

Hcompany Releases Holo4 Agentic Models

Hcompany has released Holo4, a new series of agentic models designed to operate across multiple computer interfaces, including GUIs, code, and APIs. The release directly addresses a common limitation in AI agents, which are often specialized for a single type of interaction. By creating a single model that can fluidly switch between clicking on a screen, writing code, or making API calls, Hcompany is targeting the complexity of real-world business workflows that are not confined to a single application environment.

Technical Specifications and Performance

The Holo4 series is built on Qwen foundation models and introduced in two main sizes, alongside a smaller model developed in partnership with NVIDIA. While Holo4 doesn't top the leaderboards against significantly larger closed models like Opus 5.5, its primary value proposition is delivering competitive performance at a much lower parameter count and operational cost. The company has also open-sourced the model weights and performance trajectories to promote transparency and further development.

• Models Released: Holo4-27B (dense), Holo4-35B-A3B (Mixture-of-Experts), and Holotron4 Nano (based on NVIDIA Nemotron 3).
• Multi-Interface Operation: The same model can operate on desktops, web browsers, Android, and in code sandboxes.
• Training Data: Trained on a proprietary set of ~10,000 tasks generated by an internal 'Agentic Task Factory'.
• OSWorld 2.0 Benchmark: The 27B model achieves a 61.7% success rate, demonstrating strong capability for its size.

The release of Holotron4 Nano, derived from applying Hcompany's post-training methods to an NVIDIA Nemotron model, indicates that their agent-tuning process is model-agnostic and can be used to enhance other generalist foundation models. This suggests a repeatable strategy for creating specialized agents without building from scratch, potentially accelerating the development of capable, efficient models for specific enterprise needs.

Strategic Takeaway: Hcompany is positioning Holo4 not as a benchmark chaser but as a pragmatic tool for enterprise automation, betting that cost-efficiency and multi-interface flexibility are more critical for real-world adoption than raw performance against the largest proprietary models.