Holo4 Open-Weight Agents: Screens, Code and MCP in One

Holo4 is H Company's new family of open-weight agent models, released on September 28, 2026, and its pitch is simple: one model that works through whatever interface a task offers. Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools, choosing whichever fits. It arrives in two sizes on Hugging Face and on H Company's own API.

• Holo4-27B is a dense model; Holo4-35B-A3B is a mixture-of-experts model.
• Weights are published in BF16, FP8, NVFP4 and 4-bit GGUF.
• H Company reports 61.7% on OSWorld 2.0 for Holo4-27B.
• A separate Holotron4 Nano (30B-A3B) builds on NVIDIA's Nemotron 3 Nano Omni.

What is Holo4, and how is it different from Holo3.1?

Our coverage of Holo3.1 focused on fast, private computer-use agents that operate a graphical interface. Holo4 widens the scope. Instead of treating screen control, coding and tool calls as separate specialties, H Company trained one model to move between desktops, web apps, Android, code sandboxes and business APIs.

That matters because real workflows are mixed. A task might start in a web dashboard, need a quick script to reshape a spreadsheet, then finish with an API call. A generalist computer-use agent that can pick the most direct route should waste fewer steps than one forced to click through every screen.

How was Holo4 trained?

According to H Company, Holo4 went through supervised fine-tuning on 127 billion tokens, followed by reinforcement learning with two expert models that were merged. The training tasks came partly from the company's Agentic Task Factory, which has generated roughly 10,000 tasks across web apps, MCP servers and desktop environments, including hybrid setups that expose the same state through both a GUI and MCP. H Company also published a trajectory viewer and dataset so researchers can inspect agent runs.

Holo4 benchmark results on OSWorld 2.0

On OSWorld 2.0, a benchmark of long desktop workflows, H Company reports 61.7% for Holo4-27B. For context, the company lists Claude Opus 5.5 at 81.8% on the same test, so the frontier closed model still leads; our Claude Opus 5.5 benchmark breakdown covers that model. The smaller mixture-of-experts model scored 30.9% on OSWorld 2.0. H Company's argument is about efficiency: on AutomationBench it says Holo4 competes well on cost per task. These are the developer's own figures and have not yet been independently reproduced.

Can you run Holo4 locally?

The format spread is the practical headline for local AI builders. FP8 and NVFP4 target modern GPUs, while the 4-bit GGUF release fits the llama.cpp ecosystem. The 35B-A3B mixture-of-experts design activates only about 3 billion parameters per token, which generally makes it lighter to serve than its total size suggests. H Company did not state a license in its announcement, so check the model card before commercial use.

Open weights for a generalist agent give developers something to study, fine-tune and run on their own infrastructure. Follow our AI coverage for independent test results as they appear.

Sources: H Company on Hugging Face: Holo4, powering generalist computer-use agents — September 28, 2026; HyperAI: Holo4 launches generalist agentic models for multi-interface workflows — September 28, 2026. Benchmark and comparison figures are H Company's own.