H Company releases Holo4, open-weight models for agents that operate a computer

H Company, a French AI lab, released Holo4 on September 28, 2026: a pair of models built to drive software on their own. Holo4 can click and type on a screen, write and run its own code, and call MCP or API tools, and the company says it picks whichever fits the task. The weights for both sizes are on Hugging Face, and both are also served through the company's H Models API.

What was released

The two models differ in architecture as well as size. Holo4-27B is a dense model with 27 billion parameters, built on Qwen3.8. Holo4-35B-A3B is a mixture-of-experts model with 35 billion parameters in total, built on Qwen3.5 MoE. Both configurations allow a context of up to 262,144 tokens.

H Company says the same model runs on desktops, on the web, on Android, in a code sandbox, and against business APIs, and that it is called the same way everywhere. Next to Holo4, the company released Holotron4 Nano, a smaller agent model built on Nvidia's Nemotron 3 Nano Omni.

• Holo4 weights come in BF16, FP8, NVFP4, and 4-bit GGUF, so the quantized files can go straight into local runtimes that read GGUF.
• Holotron4 Nano ships only in BF16 and FP8, under the NVIDIA Open Model Agreement.
• Every agent trajectory behind the published benchmark scores is available as a dataset on Hugging Face.
• The model cards point to the hai-agents harness, which sends screenshots and tool results to the model and carries out the actions it requests.

What the vendor's numbers show

According to H Company's measurement, Holo4-27B scores 61.7% on OSWorld 2.0, a benchmark for controlling a desktop. Holo4-35B-A3B reaches 30.9%. In the same chart the company lists Opus 5.5 at 81.8%. On AutomationBench, which tests work through APIs, the company reports 45.4% for the larger model and 34.5% for the smaller one.

The company estimates that an OSWorld 2.0 task costs $1.22 with Holo4-27B and $0.61 with Holo4-35B-A3B. Those costs are calculated from the tokens of each run at H Models API rates. H Company itself notes that benchmark releases, task subsets, and harnesses differ between the models it compares, and no independent measurement has been published yet.

The license decides which model you can ship

The two sizes are not licensed the same way. Holo4-27B is released under CC BY-NC 4.0, which does not allow commercial use. Holo4-35B-A3B is released under Apache 2.0, which does.

This suggests a trade-off for anyone building a product. The model you can put into a paid service without a separate agreement is the one that, by the company's own numbers, scores roughly half as much on OSWorld 2.0. The announcement does not list a per-token price for the H Models API.