Today we're releasing Holo4, our new family of generalist computer-...
H@hcompany_ai
20 views
Oct 01, 2026
~1 min read
Advertisement
1
Today we're releasing Holo4, our new family of generalist computer-use models.
Holo4 can click, code, and call tools across desktop, web, Android, code sandboxes, MCP servers, and business APIs.
Available today on the H Models API for commercial use.
Holo4 can click, code, and call tools across desktop, web, Android, code sandboxes, MCP servers, and business APIs.
Available today on the H Models API for commercial use.
2
Holo4 ships in two variants:
- Holo4 27B, a dense model
- Holo4 35B-A3B, a Mixture-of-Experts model
Find them on:
- H Model API: hcompany.ai/models-api
- H’s HuggingFace Collection with weights in BF16, FP8, NVFP4, and 4-bit GGUF: huggingface.co/collections/Hc…
- Holo4 27B, a dense model
- Holo4 35B-A3B, a Mixture-of-Experts model
Find them on:
- H Model API: hcompany.ai/models-api
- H’s HuggingFace Collection with weights in BF16, FP8, NVFP4, and 4-bit GGUF: huggingface.co/collections/Hc…
3
Holo4 27B delivers strong results across computer-use evaluations:
• 85.2% on OSWorld
• 61.7% on OSWorld 2.0
• 44.1% on ALE-CLI
• 85.1% on AndroidWorld
• 45.5% on AutomationBench
On AutomationBench, Holo4 27B reaches 45.4% at approximately $0.05 per task in our cost accounting.
• 85.2% on OSWorld
• 61.7% on OSWorld 2.0
• 44.1% on ALE-CLI
• 85.1% on AndroidWorld
• 45.5% on AutomationBench
On AutomationBench, Holo4 27B reaches 45.4% at approximately $0.05 per task in our cost accounting.
4
On long-horizon desktop workflows, Holo4 27B scores 61.7% on OSWorld 2.0, compared with 60.5% for GPT-6 Sol in the comparison shown.
5
We trained Holo4 on 127B tokens, including successful agentic trajectories from desktop, web, MCP, API, and mobile environments.
Our Agentic Task Factory has generated approximately 10,000 tasks across web apps, MCP servers, desktop environments, and hybrid GUI/MCP environments.
Our Agentic Task Factory has generated approximately 10,000 tasks across web apps, MCP servers, desktop environments, and hybrid GUI/MCP environments.
6
We also rebuilt the agent harness around long-running work.
Holo4 agents can use reliable memory across hundreds of steps and access a shell directly on the desktop machine.
Holo4 agents can use reliable memory across hundreds of steps and access a shell directly on the desktop machine.
7
The training recipe combines supervised fine-tuning with two asynchronous online-RL experts, specialized for different classes of workflows and merged back into one generalist model.
8
Finally, we are releasing the trajectories behind our public benchmark results so they can be inspected and replayed:
- Replay them at trajectories.hcompany.ai
- Download them from huggingface.co/datasets/Hcomp…
- Replay them at trajectories.hcompany.ai
- Download them from huggingface.co/datasets/Hcomp…