H Company has released Holo4, a new family of agentic models intended to interact with digital environments across multiple platforms. The models are designed to operate on desktops, the web, and Android devices, as well as within code sandboxes and against business APIs. This unified approach means a single model can be used for diverse tasks without needing to switch between specialized agents for each interface.
Holo4 comes in two sizes: a 27-billion parameter dense model and a 35-billion parameter Mixture of Experts (MoE) model. Both models are built upon Qwen bases, with the dense version improving upon Qwen3.8 27B. The company states that Holo4 models offer significant improvements over their Qwen base. While Holo4 trails leading closed models on long workflows, scoring 61.7% on OSWorld 2.0 with the 27B model compared to Opus 5.5's 81.8%, it achieves this with substantially fewer parameters and at a lower cost. The 35B-A3B MoE model achieved a score of 30.9% on the same benchmark.
The development of Holo4 involved an "agentic task factory" that generates interactive environments and verifiable tasks from documentation alone. This process has produced approximately 10,000 tasks across web applications, server environments, and desktop applications, including hybrid environments that expose state through both graphical user interfaces (GUIs) and server protocols. The training process included supervised fine-tuning on 127 billion tokens and reinforcement learning. A key development was the rebuilding of the execution harness, which manages the agent's context over hundreds of steps. This improved harness now includes a reliable memory system capable of tracking extensive sequences of actions and a shell for direct interaction with the desktop machine.
Holo4 is designed to be interface-agnostic, capable of performing actions such as clicking and typing on screens, writing and executing code, and calling application programming interfaces (APIs) or other tools. This contrasts with many existing agentic models that are trained for a single interface, limiting their utility when tasks require a combination of approaches.
The company is making Holo4 models available through its H Models API and as open checkpoints on Hugging Face. The weights are offered in various formats, including BF16, FP8, NVFP4, and 4-bit GGUF. An updated version of Holotron 3, named Holotron4 Nano, is also being released. Optimized DSpark drafter checkpoints are expected soon to further accelerate inference.
Holo4-27B is described as a flagship-class reasoning model for complex, multi-step tasks across web, desktop, and mobile, featuring an extended context window of 262,144 tokens. Its weights are licensed for research-only use. The Holo4-35B-A3B model is presented as a next-generation MoE model with a similar context window, offering near-flagship accuracy at a lower cost and latency. This version is available under the Apache 2.0 license, allowing for commercial use.
For users wishing to experience a computer-use agent without immediate development, H Company offers HoloTab, a free Chrome extension. HoloTab, built on earlier Holo models like Holo3-35B-A3B, allows users to interact with websites and perform tasks directly through the browser UI, such as replying to emails or researching profiles on LinkedIn.