Most large language models are strong at answering prompts but weak at navigating the desktop. H, a French AI developer, introduced Holo 4, a family of computer-use models designed to handle graphical user interfaces by pointing, clicking, scrolling, and typing. The models are built on Alibaba's Qwen 3.8 27B and Qwen 3.6 35B-A3B and fine-tuned with supervised training and reinforcement learning.
According to The Register, H says the models cover all three computer-use categories — command lines, APIs, and GUIs — and outperform significantly larger frontier models from OpenAI on the OSWorld 2.0 benchmark. The caveat: they often cost more per task because they generate substantially more 'thinking' tokens to reach a result. The Register advises treating vendor benchmarks with skepticism.
The open-weights models are small enough for modest hardware. A 24GB Nvidia RTX 3090 should run them at 4-bit precision, and H offers BF16, FP8, NVFP4, and GGUF weights for Llama.cpp, LM Studio, and Ollama. The company also plans DSpark draft weights for speculative decoding and has released an open-source HAI-Agents harness on GitHub. H is not alone in this space: AWS, OpenAI, Google, and Anthropic are also investing in computer-use models.