H Company releases a Qwen-architecture-based Computer Use model Holo4

Article

Holo4 Release: a Qwen base specifically enhanced for Computer Use Beating AI news flash: French AI company H Company has released Computer Use models Holo4, including 27B and 35B-A3B two versions. The former is based on a Qwen3.8 dense architecture, while the latter continues post-training based on a Qwen3.5 MoE architecture.

H Company primarily enhanced computer operation and long-task execution capabilities, further training the original general-purpose model into an Agent model for Computer Use.

During supervised fine-tuning, H Company used a total of 127000000000 tokens, approximately three-quarters of which were successful Agent operation trajectories covering desktops, webpages, mobile phones,MCP and API. It then conducted reinforcement learning, separately training two types of expert capabilities: “desktop and webpages” and “terminals,MCP and API”, before merging them back into the same model. The company also generated approximately 10000 software-operation tasks for training cross-application and long-process workflows.

H Company also redesigned the Agent harness, enabling the model to continuously remember earlier operations across tasks involving hundreds of steps and to directly call the shell on a computer.

The results are mainly reflected in long tasks.Holo4 27B on OSWorld 2.0 achieved 61.7%，Qwen3.8-27B with an official score of 48.0%; on standard OSWorld,Holo4 was 85.2%，Qwen3.8-27B while 84.3%was harness. However, the https://m.theblockbeats.info/flash/369556

Key points

French AI company H Company released Computer Use models Holo4, including 27B and 35B-A3B two versions.
Holo4 27B Based on Qwen3.8 dense architecture, 35B-A3B is based on Qwen3.5 MoE architecture and underwent further post-training.
The model's training covered desktop, web, mobile,MCP and API scenarios, while strengthening long-task execution capabilities.
Holo4 27B scores, task versions, and evaluation settings are not fully consistent, so they cannot be treated as a strict comparison under identical conditions. Original link 2.0 was obtained on OSWorld 61.7%and on standard OSWorld 85.2%。
The original text points out that the relevant evaluations’harness, task versions, and evaluation setups are not entirely consistent.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Holo4 is a model specifically designed to operate computers, rather than merely answer questions. It has been trained on desktop, web, mobile, terminal,MCP and API tasks, and can also remember previous actions across tasks involving hundreds of steps and directly invoke the shell on a computer.

Why it matters to readers

This shows that general-purpose models can undergo post-training to further adapt to cross-application, long-process Computer Use Agent tasks.

If you understand the model as a software assistant that operates computers,Holo4 the focus is on having it complete more consecutive steps, rather than execute only a single instruction.

Risks and unknowns

The material comes from a media news brief and does not provide an official release, weights, or technical report from H Company for independent confirmation.
Holo4 and Qwen The evaluation conditions are not fully consistent with those of the model, so no strict performance advantage under identical conditions can be inferred from this.
The material does not state Holo4 whether it is available for download or provides API or supports commercial deployment.

H Company releases a Qwen-architecture-based Computer Use model Holo4

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center