Holo4: powering generalist computer-use agents

Hugging Face Blog · Published Sep 28, 2026

Holo4 is a series of agentic models comprising a 27B dense model and a 35B-A3B Mixture of Experts model, both accessible via the H Models API, alongside an updated Holotron4 Nano. It interacts with software through GUIs, code, MCP, and APIs, selecting the most suitable interface per task rather than being constrained to one. Its training data comes from an internal Agentic Task Factory that generates approximately 10,000 verifiable tasks from documentation—including screenshots and open-source software—across web apps, MCP servers, desktop environments, and hybrid GUI/MCP setups. Read 3 viewpoints with supporting evidence and source links.

Maxime Theillard, Frederic Renard, Vincent Coyette, Emrick Sinitambirivoutin, Avshalom Manevich, Antonio Loison, Antoine Bonnet, Maxime Langevin, Aleix Cambray (H-AI), Léonard Benedetti, Tony Wu, Mats L. Richter, Michael Eickenberg, Sławek Mucha, Matthias Brunel, Daniel Beechey

Read key viewpoints Open Publisher

Understand this piece

3 key points

Synthesis

Two-size agentic model series
Holo4 is released in two sizes: a 27B dense model and a 35B-A3B Mixture of Experts model, both available via the H Models API, alongside an updated Holotron4 Nano.
Supporting evidence 1
Original excerpt
Holo4 is our new series of agentic models. It comes in two sizes: 27B dense and 35B-A3B Mixture of Experts. Both are available on the H Models API. We are also releasing an updated version of Holotron 3: Holotron4 Nano.
Maxime Theillard, Frederic Renard, Vincent Coyette, Emrick Sinitambirivoutin, Avshalom Manevich, Antonio Loison, Antoine Bonnet, Maxime Langevin, Aleix Cambray (H-AI), Léonard Benedetti, Tony Wu, Mats L. Richter, Michael Eickenberg, Sławek Mucha, Matthias Brunel, Daniel Beechey · Paragraph 1
Holo4: powering generalist computer-use agents · Sep 28, 2026
Read in source context →
Continue exploring
model architecture →
Unified model for GUI, code, MCP, and API interfaces
Holo4 interacts with software through any available interface—GUIs, code, MCP, and APIs—and selects the most suitable interface per task, unlike most agentic models trained for only one interface type.
Supporting evidence 1
Original excerpt
Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools. It uses whichever fits the task. Most agentic models are trained for one interface only: GUI-focused models are blind without a screen, while models that prefer tool calling are stuck in front of an application that has no API. Real work is not siloed that way, and a single business task can require combining these different approaches.
Maxime Theillard, Frederic Renard, Vincent Coyette, Emrick Sinitambirivoutin, Avshalom Manevich, Antonio Loison, Antoine Bonnet, Maxime Langevin, Aleix Cambray (H-AI), Léonard Benedetti, Tony Wu, Mats L. Richter, Michael Eickenberg, Sławek Mucha, Matthias Brunel, Daniel Beechey · Paragraph 9
Holo4: powering generalist computer-use agents · Sep 28, 2026
Read in source context →
Continue exploring
interface generalization →
Agentic Task Factory produces ~10,000 verifiable tasks
The internal Agentic Task Factory generates interactive environments and verifiable tasks from documentation alone—including screenshots and open-source software—and has produced about 10,000 tasks across web apps, MCP servers, desktop environments, and hybrid GUI/MCP environments.
Supporting evidence 1
Original excerpt
Our internal set of agentic pipelines builds interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open-source software. So far it has produced about 10,000 tasks across web apps, MCP servers and desktop environments, including hybrid environments that expose the same state through a GUI and MCP.
Maxime Theillard, Frederic Renard, Vincent Coyette, Emrick Sinitambirivoutin, Avshalom Manevich, Antonio Loison, Antoine Bonnet, Maxime Langevin, Aleix Cambray (H-AI), Léonard Benedetti, Tony Wu, Mats L. Richter, Michael Eickenberg, Sławek Mucha, Matthias Brunel, Daniel Beechey · Paragraph 30
Holo4: powering generalist computer-use agents · Sep 28, 2026
Read in source context →
Continue exploring
training data generation →

Key passages3

Attributed passages with the context to verify them. Open the original text to check the source.

Paragraph 30

training data generation

Agentic Task Factory produces ~10,000 verifiable tasks

Original excerpt

Our internal set of agentic pipelines builds interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open-source software. So far it has produced about 10,000 tasks across web apps, MCP servers and desktop environments, including hybrid environments that expose the same state through a GUI and MCP.

Paragraph 1

model architecture

Two-size agentic model series

Original excerpt

Holo4 is our new series of agentic models. It comes in two sizes: 27B dense and 35B-A3B Mixture of Experts. Both are available on the H Models API. We are also releasing an updated version of Holotron 3: Holotron4 Nano.

Paragraph 9

interface generalization

Unified model for GUI, code, MCP, and API interfaces

Original excerpt

Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools. It uses whichever fits the task. Most agentic models are trained for one interface only: GUI-focused models are blind without a screen, while models that prefer tool calling are stuck in front of an application that has no API. Real work is not siloed that way, and a single business task can require combining these different approaches.

Mentioned here

All mentioned things

Holo4

Supports

The authors support Holo4, presenting it as a new series of agentic models available in two sizes (27B dense and 35B-A3B MoE) via the H Models API.

Read supporting evidence · Maxime Theillard, Frederic Renard, Vincent Coyette, Emrick Sinitambirivoutin, Avshalom Manevich, Antonio Loison, Antoine Bonnet, Maxime Langevin, Aleix Cambray (H-AI), Léonard Benedetti, Tony Wu, Mats L. Richter, Michael Eickenberg, Sławek Mucha, Matthias Brunel, Daniel Beechey

Source & methodology

These viewpoints are linked to their original sources. Paraphrases are labeled and are not verbatim quotes.

Open transcript or source material Report an issue