AI company H Company released Holo4, a family of open models built to operate software across graphical interfaces, code and APIs from one checkpoint, according to a post on Hugging Face's blog.
The family ships in three sizes: Holo4-27B, a dense model; Holo4-35B-A3B, a mixture-of-experts model; and Holotron4 Nano, a 30B-A3B model built on Nvidia's Nemotron 3. All three are available through the H Models API and on Hugging Face in multiple formats, including BF16, FP8, NVFP4 and 4-bit GGUF, the company said.
On OSWorld 2.0, a benchmark that scores how well a model completes tasks inside a real desktop environment, Holo4-27B reached 61.7%, behind Claude Opus 5.5's 81.8%, but at a fraction of the cost per task, and the smaller model held its own against frontier systems on AutomationBench, according to H Company.
The models were trained with supervised fine-tuning on 127 billion tokens, then reinforcement learning against roughly 10,000 tasks drawn from an internal system H Company calls its Agentic Task Factory, spanning desktop, web, Android, code-sandbox and business-API environments, the company said.
Computer-use agents have mostly meant pointing a frontier model at a screenshot and hoping it clicks the right pixel. A model built to work the same way across a browser, a terminal and an API call, at a fraction of a frontier model's cost, is a bet that most agent tasks need reliable execution more than they need frontier-grade reasoning.