H Company Releases Holo4: Open-Weight Computer-Use Models That Click, Code and Call Tools Across Desktop, Web, Android and APIs
MarkTechPost Michal Sutter ● Covered by 2 sources
H Company just released Holo4, a model family that can click, type, code and call tools across desktop, web, Android and APIs. One version is open for commercial self-hosting; the other isn’t, and the benchmarks show cheaper agent runs than some rivals.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
H Company has put out Holo4, a pair of computer-use models built for AI agents that need to act on real screens instead of just chatting about them. The models can click, type, write code and call MCP or API tools, and H says the same setup works across desktop, web, Android, code sandboxes and business APIs.
There are two versions. Holo4 27B is a dense model fine-tuned from Qwen3.8-27B. Holo4 35B-A3B is a mixture-of-experts model built on Qwen3.6-35B-A3B with 3B active parameters. Both are offered through H Models API with a 256K context window, but the licensing split matters: 35B-A3B comes with Apache 2.0 weights for commercial self-hosting, while 27B is CC BY-NC 4.0 and commercial use goes through the API.
H Company’s pitch is that agents usually fail in two different ways. GUI-only systems need a screen. Tool-calling systems get stuck when an app has no API. Holo4 is meant to bridge that gap by pairing the model with H’s open hai-agents harness, which feeds screenshots and tool results back into the model and then executes the requested actions.
The company is leaning hard on cost. In its table, Holo4 27B scores 85.2% on OSWorld at $0.08 per task, compared with 84.3% for its Qwen3.8-27B base at $0.22. On AndroidWorld, it reaches 85.1%. But the longer workflows still expose gaps: on OSWorld 2.0, Holo4 27B gets 61.7% at $1.22 per task, while Claude Opus 5.5 is listed at 81.8% at $8.48.
The training pipeline is very much an agent factory. H says it built about 10,000 tasks from documentation, screenshots and real software, and only kept tasks that could be verified through the live interface. Its supervised set totals 127B tokens, with roughly three quarters coming from successful agent trajectories. It also trained two LoRA experts with asynchronous online RL, one for desktop and web, the other for terminal, MCP and API work, then merged them back together.
H Company also released Holotron4 Nano, built on NVIDIA’s Nemotron 3 Nano Omni through the Nemotron Coalition. The company says that setup lifts OSWorld from 21.0% to 76.3% over the base model. Holo4 itself is available with an OpenAI-compatible API at api.hcompany.ai/v1, and the weights on Hugging Face come in BF16, FP8, NVFP4 and 4-bit GGUF formats.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI release: less cosplay, more plumbing. The market has had enough models that can talk about doing work; the interesting part is one that can actually survive a messy desktop and an API-less app without falling over. H Company is also quietly making a larger point: open weights plus ugly, real-world tasks beats glossy demo theater every time.
Read more about this at: MarkTechPost