Holo4: powering generalist computer-use agents
2026-09-28 · Hugging Face
Holo4: Powering Generalist Computer-Use Agents
On September 28, 2026, the Hcompany team released Holo4, a new series of agentic models. It comes in two sizes: a 27B dense model and a 35B-A3B Mixture of Experts model, both available on the H Models API. An updated version of Holotron 3, Holotron4 Nano, was also released.
Models Built for Every Interface
Holo4 builds on previous models and interacts with software through any available interface:
- GUIs: Clicks and types on a screen.
- Code: Writes and runs its own code.
- MCP and APIs: Calls MCP or API tools.
Most agentic models are trained for a single interface only. GUI-focused models are blind without a screen, while models preferring tool calling are stuck with applications lacking an API. Holo4 uses whichever interface fits the task, combining different approaches as real business workflows require.
Cross-Platform Unified Experience
Holo4 runs across multiple platforms:
- Desktops
- The Web
- Android
- Code sandboxes
- Business APIs
In every case, it is the same model and is called the same way, eliminating the need to select different models for different platforms.
Performance and Cost-Effectiveness
Holo4 models show significant improvements over their Qwen base. On long workflows, Holo4 trails only the strongest closed models:
- On OSWorld 2.0, Holo4 27B scores 61.7% compared to 81.8% for Opus 5.5, and Holo4 35B-A3B reaches 30.9%.
- While slightly behind top closed models, Holo4 achieves this with orders of magnitude fewer parameters and at a much lower cost.
On the hardest academic benchmarks for desktop control (OSWorld 2.0) and API use (AutomationBench), Holo4 competes with frontier models at a fraction of the cost per task. The team has open-sourced every trajectory behind their public benchmark scores, which can be replayed online or downloaded from Hugging Face.
AI That Does Work
Trained on environments and tasks from the Agentic Task Factory, Holo4 models excel at professional software. The team demonstrated examples using Holo4 27B alongside its base model, Qwen3.8 27B, with the same prompts and harness:
3D Modeling: Eiffel Tower
The task involved building a 3D model of the Eiffel Tower in FreeCAD according to specific design requirements, such as a square plan at every height and smooth curved legs. Holo4 27B completed the task using 84 calls and 1.3 million tokens.
3D Modeling: H Logo
The task was to build a 3D model of the H company logo in FreeCAD, consisting of a solid filled disc beside a blocky sans-serif capital letter H. Holo4 27B completed it using 94 calls and 1.5 million tokens.
Game Design: Pac-Man
The task required building a Pac-Man-style game in the Godot engine and leaving it running, involving laying out a rectangular maze of walls on a grid with pellets.