What can Holo4 do that other agent models cannot?

Hcompany has released Holo4, a new series of agentic models that operate across graphical interfaces, code sandboxes, MCP servers and business APIs without switching models. The two main sizes — 27B dense and 35B-A3B mixture of experts — are already served through the H Models API, and the weights sit on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF.

One model for every interface

Most agentic models are trained for a single interface. A GUI-focused model goes blind without a screen. A tool-calling model stalls in front of an application that exposes no API. Holo4 clicks and types on a screen. It writes and runs its own code. It also calls MCP or API tools, choosing whichever fits the task. The model runs on desktops and on the web. It works on Android and in a code sandbox. It also connects to business APIs. The same weights are called the same way in each case. You do not need to select a different model for each platform.

Benchmarks and cost

On OSWorld 2.0, Holo4 27B scores 61.7% against 81.8% for Opus 5.5. Holo4 35B-A3B reaches 30.9%. The company says it achieves these results with orders of magnitude fewer parameters and at a much lower cost. Costs on the OSWorld chart are estimated from the input and output tokens of each agentic run. Holo4 is priced at H Models API rates for a single run. Qwen3.8 27B uses the model card score and cost from the tokens of Hcompany's run at Alibaba Cloud list prices. Qwen3.6 35B-A3B uses a single run in their harness at Alibaba Cloud list prices with cache hits at 20% of the input price. OpenAI launch data supplies the GPT and Opus effort sweeps. Other closed and open-weight points use the official OSWorld 2.0 leaderboard. Releases differ. Harnesses differ. Task subsets differ.

On AutomationBench v1.0.6, Holo4, Qwen3.8 27B and Qwen3.6 35B-A3B scores and costs were measured in Hcompany's internal harness. Other models use public-set scores from the AutomationBench README. Cost per task comes from the official leaderboard, which runs on the private set. Hcompany will report Holo4 on the private set once it is evaluated.

Training factory and harness

The Agentic Task Factory builds interactive environments and verifiable tasks from documentation alone. It uses screenshots of real websites or open-source software. So far it has produced about 10,000 tasks across web apps and MCP servers. It also covers desktop environments, including hybrid environments that expose the same state through a GUI and MCP. Alongside training, the team rebuilt the harness. The harness is the loop that executes the model's actions and manages its context over hundreds of steps. The largest changes were giving the agent a reliable memory that can keep track of hundreds of steps. Another change was adding a shell on the desktop machine itself. Agents tagged why each task failed. Engineers reviewed their fixes.

Holotron4 Nano

As a member of the NVIDIA Nemotron Coalition, Hcompany applied its latest post-training stack to the Nemotron 3 Nano Omni model. The same recipe turns Nemotron 3 Nano Omni into Holotron4 Nano. This is a generalist agentic model that significantly improves over the base model on GUI workflows. It also improves in environments exposing MCP, APIs or coding sandboxes. Gains are absolute percentage-point improvements over Nemotron 3 Nano Omni. Nothing in the recipe is size-specific.

What ships today

Both Holo4 sizes are available today on the H Models API. Weights are on Hugging Face in BF16 and FP8. They are also available in NVFP4 and 4-bit GGUF, next to Holotron4 Nano. Optimized DSpark drafter checkpoints will follow in the coming days to further accelerate inference. Every trajectory behind the public benchmark scores is open for replay at trajectories.hcompany.ai. The trajectories are also available for download from Hugging Face.