Summary
Sam Witteveen examines Holo4, a new open-weight model from Paris-based H Company that is trained to click through graphical interfaces, write and run code, and call tools through MCP servers and APIs. The aim is to replace the common practice of chaining separate visual and orchestration models.
The video starts with the agent loop, in which the model only decides and the harness does the acting, and explains the three action lanes Holo4 targets: GUIs, code, and tools. It then covers the model family, including a 27B dense version and a mixture-of-experts version, built on Qwen. The full BF16 weights are 54 GB, and he runs them locally on a Dell Pro Max with an RTX Pro 6000, noting that screenshots fill the KV cache quickly.
Benchmark results are compared against recent frontier models. On OSWorld, the 27B model sits close to them and slightly ahead of GPT-5.5, at roughly one-seventh the cost of Opus 5.5 and Astra. The dense model also clearly outperforms the mixture-of-experts version on visual tasks. Finally, he explains H Company’s agentic task factory, which builds automatically checkable environments to teach the model which action type to choose.
📺 Source: Sam Witteveen · Published October 06, 2026
🏷️ Format: Deep Dive
