# Holo4: A Model That Clicks, Codes and Calls Tools **URL:** \ **Category:** Content Creators\ **Tags:** sam-witteveen, function-calling, agents, machine-learning, tools, coding\ **Created:** [6 October 2026 14:10 UTC](https://www.artofsm.art/t/holo4-a-model-that-clicks-codes-and-calls-tools/25127 "2026-10-06T14:10:32Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [6 October 2026 14:10 UTC](https://www.artofsm.art/t/holo4-a-model-that-clicks-codes-and-calls-tools/25127/1 "2026-10-06T14:10:32Z")
[![](https://www.artofsm.art/uploads/default/original/3X/8/f/8fd28d0a754ceee41b4d403b93ae39ea95ae7aab.jpeg "Holo4: A Model That Clicks, Codes and Calls Tools") ](https://www.youtube.com/watch?v=vKjnZAcimss) Holo4, developed by H Company, is a unified AI model that dynamically integrates GUI interactions, code execution, and API/tool calls to handle complex, multi-modal tasks within a single framework, offering flexibility and efficiency across diverse workflows. Trained through supervised fine-tuning and reinforcement learning, Holo4 delivers competitive performance with cost-effective, locally runnable models, though it is released under a non-commercial license with resources available for research and experimentation. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [6 October 2026 14:31 UTC](https://www.artofsm.art/t/holo4-a-model-that-clicks-codes-and-calls-tools/25127/3 "2026-10-06T14:31:04Z")
The video discusses Holo4, a model developed by H Company in Paris, designed to seamlessly integrate three key capabilities for AI agents: interacting with graphical user interfaces (GUIs) through clicking and typing, writing and executing code, and calling APIs or tools (MCPs). Traditionally, models have specialized in either visual interaction or API/tool usage, but real-world tasks often require a combination of these skills. Holo4 stands out by being trained to choose dynamically among these three “lanes”—GUI, code, and tools—allowing it to handle complex workflows such as navigating legacy UIs, processing data with code, and posting results via APIs, all within a single unified model. The agent operates in a loop where it receives a task, decides on an action (click, code, or tool call), and then a harness executes that action, feeding the results back to the model along with memory of previous steps. This separation between decision-making and execution is crucial, especially as the model’s performance depends heavily on the harness that runs it. Holo4 supports three main action types: GUI interactions via screenshots, code execution in a sandbox, and calling structured APIs or tools. Each has its strengths and limitations, with GUI being flexible but slower and prone to breaking, code being fast and precise but limited to programmable tasks, and tools being reliable but dependent on available APIs. H Company released two main versions of Holo4: a 27-billion parameter dense model and a 35-billion parameter mixture-of-experts model, along with a smaller Holotron Nano variant. Benchmarks show the dense 27B model generally outperforms the mixture-of-experts model, especially in tasks involving image understanding. While Holo4’s performance is competitive with recent frontier models, it offers a more cost-effective and locally runnable alternative. The training process involves supervised fine-tuning on a large dataset followed by reinforcement learning in an “agentic task factory,” a suite of interactive environments that simulate real-world tasks combining GUI, code, and tool usage. This training encourages the model to pick the most efficient action path for each task. The video also demonstrates Holo4’s capabilities through various examples, showing how it can navigate a GUI to extract data, switch to code for processing, or use APIs when available to streamline tasks. This flexibility allows it to adapt to different environments and task requirements dynamically. The model’s ability to handle multiple modalities within a single framework is a significant advancement, enabling more practical and versatile AI agents that can operate across diverse interfaces and workflows without needing separate specialized models. Finally, the video notes that while Holo4 is open-weight and available for experimentation, it is released under a non-commercial license, limiting its use in commercial products without going through H Company’s API or direct licensing. The creators have made benchmark trajectories publicly available, providing valuable resources for researchers and developers interested in building similar multi-modal agents. Overall, Holo4 represents a promising step toward more integrated and capable AI agents that can intelligently choose between clicking, coding, and calling tools to accomplish complex tasks efficiently. ## Useful Links - [Holo4 Blog by H Company](https://hcompany.ai/newsroom/holo4) — Direct source of official information about Holo4 model and its capabilities. - [Holo4 27B Model on HuggingFace](https://huggingface.co/Hcompany/Holo4-27B) — Provides access to the main dense Holo4 model weights for local use and research. - [Holo4 35B-A3B Model on HuggingFace](https://huggingface.co/Hcompany/Holo4-35B-A3B) — Access to the alternative Holo4 model variant for experimentation and benchmarking. - [Holotron4 Nano Model on HuggingFace](https://huggingface.co/Hcompany/Holotron4-30B-A3B) — Provides a smaller model variant for local experimentation and research. - [Trajectory Viewer for Holo4](https://trajectories.hcompany.ai/) — Enables detailed analysis of Holo4’s decision-making and action sequences for research and development.