H Company has released Holo4, a family of generalist computer-use vision-language models designed to click, type, write code, and call tools across desktop, web, Android, and API environments. The release includes two model sizes: Holo4 27B (dense) and Holo4 35B-A3B (Mixture of Experts with 3B active parameters), both supporting a 256K context window.
- Holo4 35B-A3B is licensed under Apache 2.0 for commercial self-hosting, while Holo4 27B uses CC BY-NC 4.0 and requires the H Models API for commercial use.
- The models are built on Qwen3.8-27B and Qwen3.6-35B-A3B bases, paired with H's open hai-agents harness for executing actions via screenshots and tool results.
- On OSWorld, Holo4 27B scores 85.2% at $0.08 per task, compared to its base Qwen3.8-27B score of 84.3% at $0.22.
- In long workflows on OSWorld 2.0, Holo4 27B achieves 61.7% accuracy at $1.22 per task, while Claude Opus 5.5 scores 81.8% at $8.48.
- Training involved an Agentic Task Factory producing ~10,000 tasks, supervised fine-tuning on 127B tokens, and asynchronous online RL with two LoRA experts.
- H Company also released Holotron4 Nano, which lifts OSWorld scores from 21.0% to 76.3% over its Nemotron 3 Nano Omni base.
The models provide a unified interface for agentic tasks across diverse platforms, offering lower-cost alternatives to frontier proprietary models while maintaining open-weight accessibility for self-hosting or API usage.