H Company releases Holo4 generalist computer-use agent models

H Company has released Holo4, a series of agentic models designed for generalist computer-use tasks. The series includes two model sizes: a 27B dense model and a 35B-A3B Mixture of Experts model. These models are designed to interact with software through all major interfaces, including GUI, code, MCP, and API, allowing a single model to operate across desktop, web, Android, code sandboxes, and business APIs without needing platform-specific model selection.

Benchmark Performance and Cost Efficiency

Holo4 demonstrates significant performance improvements over Qwen base models. On the OSWorld 2.0 benchmark for desktop control, Holo4 27B achieved a score of 61.7%, while Holo4 35B-A3B scored 30.9%. For comparison, the closed-source Opus 5.5 scored 81.8% on the same benchmark. In AutomationBench (API usage), Holo4 shows competitive performance with frontier models. Despite these capabilities, Holo4 requires far fewer parameters and offers task costs that are orders of magnitude lower than larger closed models.

Training Methodology

The models were trained using supervised learning and reinforcement learning within an Agentic Task Factory. This internal pipeline generates large-scale environments and tasks from documents like website screenshots and open-source software, creating approximately 10,000 tasks so far. The training pipeline involved supervised fine-tuning on 127B tokens, followed by two RL experts and a merged model. The harness was rebuilt to improve action execution and context management over hundreds of steps, incorporating reliable memory and direct shell access to the desktop machine.

Holotron4 Nano and Availability

H Company also released Holotron4 Nano, an update to Holotron 3 based on the NVIDIA Nemotron 3 Nano Omni model. This conversion applies H Company's latest post-training stack to improve performance in GUI workflows, MCP, APIs, and coding sandboxes. The recipe is not size-specific and transfers well to converting generalist models into agentic experts.

All models are available via the H Models API. Weights for Holo4-27B, Holo4-35B-A3B, and Holotron4 Nano (Hcompany/Holotron4-30B-A3B) are hosted on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats. Public trajectories are available for viewing and download, and optimized DSpark drafter checkpoints for inference acceleration are expected within a few days.