Holo4 Released: Qwen Foundation, Specifically Enhanced for Computer Use

Beating AI News Flash: French AI company H Company has released the Computer Use model Holo4, including two versions: 27B and 35B-A3B. The former is based on the Qwen3.8 dense architecture, while the latter continues post-training based on the Qwen3.5 MoE architecture.

H Company mainly strengthened computer operation and long-task execution capabilities, further training the original general-purpose model into an Agent model oriented toward Computer Use.

During the supervised fine-tuning stage, H Company used a total of 127 billion tokens, about three-quarters of which were Agent operation trajectories that successfully completed tasks, covering desktop, web, mobile, MCP, and API. It then conducted reinforcement learning, separately training two types of expert capabilities, "desktop and web" and "terminal, MCP, and API," and then merged them back into the same model. The company also generated about 10,000 software operation tasks on its own to train cross-application and long-process work.

H Company also redesigned the Agent harness, allowing the model to continuously remember previous operations during tasks involving hundreds of steps and to directly call the shell on the computer.

The results are mainly reflected in long tasks. Holo4 27B scored 61.7% on OSWorld 2.0, while the official score for Qwen3.8-27B was 48.0%; on the regular OSWorld, Holo4 scored 85.2%, while Qwen3.8-27B scored 84.3%. However, the harness, task versions, and evaluation settings for these results are not completely consistent and cannot be regarded as a strict same-condition comparison.