H Company Releases Holo4 Computer-Use Models
H Company launches Holo4 vision-language models for computer use across desktop, web, Android, and APIs, scoring 85.2% on OSWorld.
Stock photo for illustration only, not from the actual event
- H Company unveils Holo4 vision-language models for cross-platform computer use
- Holo4 35B-A3B is released as open-weights under Apache 2.0 license
- Holo4 27B achieves 85.2% on OSWorld at $0.08 per task
- Supports desktop, web, Android, code sandboxes, and OpenAI-compatible APIs
H Company has officially released Holo4, a new family of vision-language models engineered specifically for computer-use automation. The models are capable of executing clicks, typing, writing code, and calling tools across multiple platforms including desktops, web browsers, Android devices, code sandboxes, and business APIs using a unified interface.
Regarding deployment options, Holo4 35B-A3B—built upon Qwen3.6-35B-A3B—ships with open weights under the Apache 2.0 license for commercial self-hosting. Meanwhile, Holo4 27B, fine-tuned from Qwen3.8-27B, operates under a CC BY-NC 4.0 license, directing commercial use through the H Models API.
Stock photo for illustration only, not from the actual event
The company targeted a persistent industry gap where GUI-only agents fail without a screen, and tool-calling agents stall when applications lack APIs. Holo4 addresses this by pairing with H's open hai-agents harness, which transmits screenshots and tool execution results directly to the model to drive automated workflows.
According to benchmark figures published by H Company, Holo4 27B scores 85.2% on OSWorld at a cost of $0.08 per task, outperforming its Qwen3.8 27B base model which scores 84.3% at $0.22. Additionally, Holo4 27B reaches 85.1% on AndroidWorld, alongside the introduction of Holotron4 Nano built on NVIDIA's Nemotron 3 Nano Omni.
The introduction of Holo4 highlights a major push toward robust general-purpose agentic execution. By leveraging a comprehensive training pipeline featuring a 10,000-task Agentic Task Factory, a 127-billion-token Supervised Fine-Tuning dataset, and asynchronous online reinforcement learning to merge LoRA experts, H Company demonstrates how modern AI infrastructure tackles the complexities of multi-step digital workflows.
To achieve this, H Company rebuilt its agent loop based on OSWorld 2.0 failure analysis, introducing reliable memory across hundreds of steps and a dedicated desktop shell. The API is fully OpenAI-compatible at https://api.hcompany.ai/v1, priced at $0.40 input and $3.00 output per 1 million tokens for the 27B model.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment