H Company Holo4: 27B Computer-Use Agent Hits 61.7% on OSWorld 2.0

Executive Overview

H Company has officially announced the launch of the Holo4 series of agentic models, available today via the H Models API with weights published on Hugging Face. The series ships in two sizes: Holo4-27B (dense) and Holo4-35B-A3B (Mixture of Experts), alongside an updated Holotron4 Nano. Holo4-27B scores 61.7% on OSWorld 2.0 — the hardest desktop-control benchmark — versus 81.8% for Opus 5.5, while H Company claims a dramatically lower cost per task. The models interact with software through any available interface: GUIs, code, MCP, and APIs, automatically selecting the best interface for each task.

Deep-Dive Features & Architecture

Holo4 is built on H Company's previous model and interacts with software through any available interface: graphical user interfaces (GUIs), code, the MCP protocol, and application programming interfaces (APIs). The model automatically selects the most suitable interface for the task — unlike most agentic models, which are trained for a single interface only. Holo4 was trained via supervised learning and reinforcement learning across a wide range of environments and tasks, including tasks generated by the company's own Agentic Task Factory, which has so far produced approximately 10,000 tasks across web applications, MCP servers, and desktop environments — including hybrid environments that present the same state through both GUI and MCP simultaneously.

The company rebuilt its harness — the loop that executes the model's actions and manages its context over hundreds of steps — based on agentic performance observations on OSWorld 2.0. Key changes include granting the agent trusted memory that tracks hundreds of steps and providing a shell on the desktop machine itself. The same post-training technique was applied to the Nemotron 3 Nano Omni model as part of H Company's membership in the NVIDIA Nemotron Coalition, producing Holotron4 Nano, which achieves absolute percentage-point improvements on GUI tasks and MCP, API, and code environments.

Benchmark & Competitive Performance

On the OSWorld 2.0 benchmark — the hardest for desktop control — Holo4-27B scores 61.7% versus 81.8% for Opus 5.5, 70.2% for Opus 5, and 66.2% for GPT-5.6 Sol. Holo4-35B-A3B scores 30.9%. H Company confirms these results are achieved with orders of magnitude fewer parameters and a far lower cost per task. On AutomationBench v1.0.6, internal measurements cover Holo4, Qwen3.8 27B, and Qwen3.6 35B-A3B.

Industry Impact & Enterprise Adoption

The Holo4 launch signals a shift toward multi-interface computer-use agents that can operate across desktop, web, Android, code sandboxes, and business APIs without requiring separate models per platform. With open weights on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats, enterprises can self-host or deploy via the H Models API. The cost-efficiency claim — orders of magnitude lower cost per task than frontier models — positions Holo4 as a practical option for large-scale desktop automation and agentic workflows, even where raw benchmark scores trail the largest proprietary models.

Conclusion

Holo4 represents a focused bet on cost-efficient, multi-interface computer-use agents. While Holo4-27B's 61.7% on OSWorld 2.0 trails Opus 5.5's 81.8%, the combination of open weights, immediate API availability, and dramatically lower per-task cost makes it a compelling option for enterprises scaling agentic automation. The updated Holotron4 Nano extends the same post-training stack to smaller deployments, broadening the addressable use cases.

Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis