H Company's Holo4: Open-Weight AI Agents Redefine Cross-Platform Automation

H Company has announced the release of Holo4, a new family of general-purpose AI models engineered to power advanced AI agents. These innovative models are designed to enable artificial intelligence to seamlessly interact with digital interfaces, from executing screen clicks and typing to generating code and invoking specialized tools or APIs.

The Holo4 series is introduced in two distinct sizes: Holo4 27B, a dense model, and Holo4 35B-A3B, which utilizes a Mixture of Experts (MoE) architecture with 3 billion active parameters. Both models provide an extensive 256K context window when accessed via the H Models API.

Flexible Deployment and Licensing

For organizations seeking commercial self-hosting capabilities, Holo4 35B-A3B weights are made available under the permissive Apache 2.0 license. In contrast, the Holo4 27B weights are licensed under CC BY-NC 4.0, indicating that commercial use of this model is facilitated exclusively through the H Models API.

Understanding Holo4's Capabilities

Holo4 is fundamentally a vision-language model optimized for computer interaction. Holo4 27B is an adaptation of the Qwen3.8-27B foundation, while Holo4 35B-A3B builds upon the Qwen3.6-35B-A3B architecture. Both models integrate with H Company’s proprietary 'hai-agents' harness, which manages the flow of information by relaying screenshots and tool outputs to the model, subsequently executing requested actions such as clicks, typing, code generation, and tool calls.

This development targets a critical void in current AI agent capabilities. Prior agents often falter without a graphical user interface or when an application lacks an API. Holo4 addresses this by operating consistently across desktops, web browsers, Android devices, code sandboxes, and business APIs, employing the same model and interaction method across all platforms.

Performance and Cost Efficiency

H Company's internal benchmarks indicate strong performance metrics for Holo4. The Holo4 27B model achieved an 85.2% score on OSWorld with an estimated cost of $0.08 per task, surpassing its Qwen3.8 27B base model, which scored 84.3% at $0.22 per task. On AndroidWorld, Holo4 27B reached an impressive 85.1%.

For more complex, extended workflows, such as OSWorld 2.0, Holo4 27B attained a 61.7% score at $1.22 per task. While this trails frontier models like Claude Opus 5.5, which scored 81.8% at $8.48 per task according to H Company’s figures, Holo4 demonstrates a significant cost advantage. On AutomationBench, Holo4 27B registered a 45.4% score for $0.05 per task, with a 49.3% score on the 120 held-out tasks. It is important to note that performance comparisons across different vendors' benchmarks should be interpreted with caution due to varying harnesses and testing methodologies. H Company maintains transparency by publishing all trajectories on its dedicated platform and Hugging Face.

The Engineering Behind Holo4

H Company employed a multi-faceted approach to develop Holo4:

• Agentic Task Factory: An internal system generates verifiable tasks from diverse sources like documentation, screenshots, and live software, ensuring robustness by rejecting near misses and requiring agents to solve challenges through real interfaces. Approximately 10,000 such tasks were created, covering web, desktop, MCP servers, and operating systems.
• Supervised Fine-tuning (SFT): A comprehensive dataset of 127 billion tokens was used for SFT, with roughly three-quarters comprising successful agentic trajectories across desktop (45%), web (14%), MCP and API (12%), and mobile (3%).
• Reinforcement Learning (RL): Two LoRA experts were asynchronously trained online—one for desktop and web environments, the other for terminal, MCP, and API interactions. These experts were then merged with equal weighting without further training.
• Enhanced Harness: The agent loop was reconstructed following an analysis of OSWorld 2.0 failures, leading to significant improvements in reliable memory retention across numerous steps and the integration of a shell for desktop environments.

Introducing Holotron4 Nano

In a related development, H Company also unveiled Holotron4 Nano, a model built in collaboration with NVIDIA's Nemotron 3 Nano Omni through the Nemotron Coalition. This joint effort reportedly boosted OSWorld performance from 21.0% to 76.3% over the base model.

Access and Pricing Information

Holo4 27B is priced at $0.40 per million input tokens and $3.00 per million output tokens. The Holo4 35B-A3B model offers a more cost-effective option at $0.30 per million input tokens and $2.00 per million output tokens. The API is compatible with OpenAI standards and is accessible at https://api.hcompany.ai/v1. Weights for local inference are available on Hugging Face in various formats, including BF16, FP8, NVFP4, and 4-bit GGUF, with documentation provided for local inference using vLLM and llama.cpp. Future updates will include DSpark drafter checkpoints for accelerated inference.