H Company has released Holo4, a suite of generalist computer-use models engineered to interact with graphical user interfaces, write code, and call APIs. The release includes Holo4 27B, a dense model fine-tuned from Qwen3.8-27B available under CC BY-NC 4.0, and Holo4 35B-A3B, a Mixture-of-Experts architecture released under the Apache 2.0 license for commercial self-hosting. Both models support a 256K context window on the company’s API platform.

To train Holo4, H Company developed an automated pipeline that generated approximately 10,000 synthetic task environments spanning desktop OS, web apps, and Model Context Protocol (MCP) servers. The training process combined a 127-billion-token supervised dataset of agentic workflows with asynchronous online reinforcement learning using two distinct LoRA experts merged into the final weights.

In OSWorld benchmarks, Holo4 27B achieved an 85.2% success rate at an execution cost of $0.08 per task. H Company also introduced Holo4 Nano, built on NVIDIA’s Nemotron 3 Nano Omni, and made all inference code, trajectory datasets, and model quantizations publicly accessible.

Why it matters

  • Provides open-weight alternatives for multi-platform computer-use agents, reducing deployment costs relative to proprietary frontier models.

  • Combines GUI interaction and structured API tool-calling in a single model, bridging performance gaps in cross-application automation.

  • Offers commercially permissive MoE weights that enable enterprise teams to run local, privacy-compliant agentic workflows.

Source: marktechpost.com