- Published on
H Company Releases Holo4: Open-Weight Computer-Use Models That Click, Code and Call Tools Across Desktop, Web, Androi...
- Authors
- Name
- DeepHorus
H Company has released Holo4, open-weight vision-language models for agents that click, type, code and call MCP or API tools. The models are Holo4 27B and Mixture-of-Experts Holo4 35B-A3B, which activates 3B parameters. Both offer a 256K context window through H Models API.
Built on Qwen3.8-27B and Qwen3.6-35B-A3B, the models use H’s hai-agents harness to interpret screenshots and tool results before acting. One model works across desktop, web, Android, code sandboxes and business APIs. Holo4 targets the gap between GUI-only agents and tool callers blocked by applications without APIs.
Holo4 35B-A3B has Apache 2.0 weights for commercial self-hosting, while 27B uses non-commercial CC BY-NC 4.0, making its API the commercial route. Pricing per million input and output tokens is $0.40/$3.00 for 27B and $0.30/$2.00 for 35B-A3B. The OpenAI-compatible API supports cloud deployment, while vLLM and llama.cpp support local use.
H reports Holo4 27B reached 85.2% on OSWorld at $0.08 per task, versus 84.3% for its Qwen base at $0.22; it also reached 85.1% on AndroidWorld. On OSWorld 2.0, it scored 61.7% at $1.22, compared with Claude Opus 5.5 at 81.8% and $8.48. H’s 45.4% AutomationBench score rose to 49.3% on held-out data, but it cautioned that different harnesses and effort levels make cross-vendor results directional.
H’s factory generated about 10,000 verifiable tasks: 4,000 web apps, 3,000 MCP servers and 3,000 desktop or OS tasks. Its 127B-token supervised set was about three-quarters successful trajectories. Two reinforcement-learning LoRA experts, for desktop/web and terminal/MCP/API work, were merged equally, and H published all trajectories.
H also released Holotron4 Nano, based on NVIDIA’s Nemotron 3 Nano Omni, raising OSWorld performance from 21.0% to 76.3%. H presents Holo4 as a lower-cost, adaptable alternative, not the benchmark leader. DSpark inference checkpoints are due shortly, but long workflows remain the central reliability test.
