H Company drops Holo1.5 for Computer Use
H Company just open-sourced Holo1.5, a family of 3B, 7B, and 72B models built specifically for Computer Use agents. The update boosts UI element localization accuracy by over 10% compared to its predecessor and handles complex software interfaces like Photoshop and VSCode. All weights are now accessible on Hugging Face.
Holo1.5 is an open-source family of foundation models built specifically for Computer Use agents, designed to help AI see screens and click elements just like a human. H Company dropped the weights for 3B, 7B, and a brand-new 72B size, all accessible on Hugging Face.
What it does
The core trick here is UI element localization, also called grounding. You give the model a screenshot and a prompt like "Open the Spotify App," and it outputs exact screen coordinates ("Click X, Y"). Holo1.5 delivers an accuracy boost of over 10% across all sizes compared to Holo1, setting a new benchmark for this specific task.
It also tackles UI Visual Question Answering (VQA). Agents cannot just click blindly; they need to understand what is on the screen to confirm their next move. Holo1.5 can answer natural-language questions like "Which tab is active?" or "Is the user signed in?" based on visual UI state, which helps the agent track context and resolve visual ambiguities.
The models were evaluated across web, mobile, and desktop environments including macOS, Ubuntu, and Windows. On ScreenSpot-Pro, a benchmark filled with high-resolution, complex professional apps like Photoshop, AutoCAD, and VSCode, Holo1.5 outperformed open weights like Qwen-2.5 VL, specialized systems like UI-TARS 1.5 and UI-Venus, and closed models like Sonnet 4.
How to try it
All model weights for the 3B, 7B, and 72B versions are open and hosted on Hugging Face. You can pull them directly into your own agent pipelines.
If you want to test prompt behavior before spinning up local compute, H Company put together an interactive prompt demo on Hugging Face Space. The team also noted that specialized tools and ready-to-run agents built on top of Holo will roll out in the coming weeks.
Why it matters
If you are building desktop automation, browser copilots, or RPA tools, Holo1.5 is an easy download to benchmark against your current stack. The 3B and 7B variants give you lightweight options for fast, local inference, while the 72B model provides serious muscle for dense GUI apps.
Relying on closed-source generalist models for Computer Use gets expensive quickly, and they often struggle with sub-pixel precision on complex professional software. We think open-sourcing dedicated localization and screen-reasoning models at this level gives developers a practical, cost-effective foundation to build actual autonomous workflows without getting locked into proprietary APIs.
Sources (1)
- hcompany.ai(opens in a new tab)Originalhcompany.ai