H Company has announced Holo4, a new set of generalist computer-use models for AI agents. These models can click, type, write code, and call tools like MCP or APIs.
Holo4 comes in two sizes: Holo4 27B and Holo4 35B-A3B. The 27B model is fine-tuned from Qwen3.8-27B, while the 35B-A3B is based on Qwen3.6-35B-A3B. Both models support a 256,000 token context window.
The 35B-A3B version ships with Apache 2.0 licensed weights for self-hosting. The 27B version uses CC BY-NC 4.0 licensed weights, so it can only be used via the H Models API.
Holo4 is designed for computer use, working on desktop, web, Android, code sandboxes, and business APIs. The same model runs on all platforms with a consistent interface.
Benchmarks show Holo4 27B performs well at a low cost. It scores 85.2% on OSWorld at $0.08 per task. It also reaches similar scores on Android and other benchmarks, with lower costs than some competitors.
The models are trained with a large dataset of about 127 billion tokens, including many agentic trajectories. They are fine-tuned with reinforcement learning techniques.
H Company also released Holotron4 Nano, built on NVIDIA’s Nemotron 3 Nano Omni, which improves OSWorld scores significantly. The models are available via an API compatible with OpenAI and on Hugging Face.
This release provides a flexible, cost-effective way to deploy AI agents across multiple platforms with consistent performance.