Holo4 Releases Qianwen Base, Specifically Enhanced for Computer Use

Insight Beating AI News Flash: French AI firm H Company has launched its Computer Use model Holo4, available in two variants: 27B and 35B-A3B. The 27B version is built on the Qwen 3.8 Dense architecture, while the 35B-A3B variant underwent post-training based on the Qwen 3.5 MoE architecture. H Company has focused on enhancing computer operation and long-task execution capabilities, refining its general-purpose model into an Agent model specifically designed for Computer Use. During the supervised fine-tuning phase, H Company utilized a total of 127 billion tokens, roughly three-quarters of which were successful Agent operation trajectories covering desktops, web, mobile, MCP, and APIs. The team then conducted reinforcement learning, training two sets of specialized capabilities separately: one for "desktops and web" and another for "terminals, MCP, and APIs", before merging both into a single model. The company also generated around 10,000 custom software operation tasks to train for cross-application and long-process workflows. H Company further rebuilt its Agent harness, allowing the model to retain memory of prior operations across hundreds of task steps and directly call the computer’s shell. The improvements are most pronounced in long-task performance. Benchmark results show Holo4 27B scored 61.7% on OSWorld 2.0, compared to Qwen 3.8-27B’s official score of 48.0%. On standard OSWorld, Holo4 achieved 85.2%, versus 84.3% for Qwen 3.8-27B. However, the harness, task versions, and evaluation settings for these results are not fully aligned, so they cannot be treated as strictly controlled, same-condition comparisons.