Holo4 combines GUI control, code, MCP and APIs in one agent model

One model for several interfaces

H introduced the Holo4 series of agentic models. It includes a dense 27-billion-parameter model and Holo4-35B-A3B, which uses a Mixture of Experts architecture. The company also updated its smaller model and released Holotron4 Nano.

Holo4 can click and type on screen, write and run code, and call MCP or API tools. The same model works on desktops, websites, Android, code sandboxes and business APIs. The goal is to let an agent switch interfaces as needed during a single workflow rather than requiring a separate model for each platform.

Results and comparison limits

In the OSWorld 2.0 evaluation, Holo4 27B scored 61.7%, while Holo4 35B-A3B reached 30.9%. H's table lists Opus 5.5 at 81.8%. The company says its models complete tasks at a lower cost, but also notes that releases, evaluation harnesses and task subsets differ across the compared systems.

The AutomationBench figures carry a similar limitation: Holo4 and Qwen were measured on the public set in H's internal harness, while figures for other models come from a leaderboard that uses the private set. H published every trajectory behind its public-benchmark scores, allowing readers to replay individual steps or download the dataset.

Training and availability

The models were trained with supervised and reinforcement learning. H says its Agentic Task Factory has produced about 10,000 tasks spanning web apps, MCP servers and desktop environments. The execution harness also gained memory for workflows lasting hundreds of steps and a shell running directly on the desktop machine.

Both Holo4 sizes are available through the H Models API. Weights published on Hugging Face come in BF16, FP8, NVFP4 and 4-bit GGUF formats. H also plans to release optimized DSpark drafter checkpoints. The reported results are company evaluations rather than independent verification.