H Company Releases Holo4: Open-Weight Agentic Models for Generalist Computer Use

H Company has announced the release of Holo4, a new series of open-weight agentic models designed for generalist computer use. These models are now available in two distinct sizes: a 27B dense model and a 35B-A3B Mixture-of-Experts (MoE) variant. Additionally, an updated Holotron4 Nano has been introduced. The model weights are accessible on Hugging Face in FP16, FP8, and GGUF formats, and can also be served via the H Models API. For broader context, explore our AI Tools Pricing.

Introducing Holo4: Agentic Models for Broad Application

H Company's Holo4 models represent a step towards more versatile AI agents capable of interacting with various computing environments. These models are specifically engineered to utilize multiple interfaces, including graphical user interfaces (GUIs), code, MCP, and API tools. Their operational scope extends across desktops, web browsers, Android devices, code sandboxes, and business APIs, aiming to provide a comprehensive solution for diverse computational tasks.

Model Architecture and Training Methodology

The Holo4 series leverages advanced training techniques to achieve its generalist capabilities. The models were developed using a combination of supervised learning and reinforcement learning. This training regimen involved a substantial collection of environments and tasks, many of which were generated by H Company's proprietary Agentic Task Factory. The base model for the Holo4 27B variant is Qwen3.8 27B, which has undergone significant improvements to form the foundation of Holo4.

Performance Benchmarks on OSWorld 2.0

In terms of performance, the Holo4 27B model achieved a score of 61.7% on the OSWorld 2.0 desktop control benchmark. H Company emphasizes transparency by open-sourcing every trajectory behind its public benchmark scores. This approach allows for each step of the model's performance to be replayed and audited, providing a verifiable record of its capabilities.

Availability and Access

The Holo4 model weights are publicly available on Hugging Face, offering flexibility for developers and researchers. Users can download the models in various formats, including FP16, FP8, and GGUF. For those seeking API-based access, the models are also served through the H Models API, facilitating integration into existing applications and workflows.

Key Takeaways for Developers and Researchers

• H Company has released Holo4, a new series of open-weight agentic models.
• Holo4 is available in 27B dense and 35B-A3B Mixture-of-Experts variants.
• Model weights are accessible on Hugging Face and via the H Models API.
• Holo4 was trained using supervised and reinforcement learning on diverse tasks.
• The models are designed for generalist computer use across multiple interfaces and environments.
• Holo4 27B scored 61.7% on the OSWorld 2.0 desktop control benchmark.
• H Company provides open-sourced benchmark trajectories for auditing.

Conclusion

The release of H Company's Holo4 models marks an expansion in the availability of open-weight agentic AI for generalist computer use. With models accessible on Hugging Face and through the H Models API, and a transparent approach to benchmarking, Holo4 aims to support a wide range of applications across various computing environments. Developers and researchers can explore these models for tasks requiring interaction with GUIs, code, and other digital interfaces.

Sources

• Holo4: powering generalist computer-use agents
• Holo4 - a Hcompany Collection
• Hcompany/Holo4-27B-GGUF · Hugging Face
• Hcompany/Holo4-35B-A3B · Hugging Face
• Hcompany/Holo4-27B · Hugging Face