Holo3 AI Model Beats GPT‑5.4 at One‑Tenth the Cost for Computer Use

H Company has launched the Holo3 series of AI models, setting a new record in the computer‑use domain.

In the OSWorld‑Verified benchmark, the Holo3‑122B‑A10B model achieved a 78.85% score, surpassing mainstream models such as GPT‑5.4 and Opus 4.6, while its operating cost is roughly one‑tenth of the competitors.

Two Versions, Different Positioning

Holo3‑122B‑A10B : flagship version offered only via API, priced at $0.40 per million input tokens and $3.00 per million output tokens.

Holo3‑35B‑A3B : lightweight version released under Apache 2.0, priced at $0.25 per million input tokens and $1.80 per million output tokens; its weights are publicly available on Hugging Face for free‑tier users.

The 35B model’s weights have already been uploaded to Hugging Face.

Designed Specifically for GUI Agents

Holo3 is a visual language model optimized for GUI agents that can operate in web, desktop, and mobile environments. It is built on the Qwen3.5 architecture and uses a sparse Mixture‑of‑Experts (MoE) design, with only 3 B active parameters in the 35B version and 10 B active parameters in the 122B version.

Enterprise‑Level Application Validation

To verify real‑world capability, H Company built the H Corporate Benchmark, which contains 486 multi‑step tasks covering e‑commerce, business software, collaboration, and multi‑application scenarios. Tests show that Holo3 performs strongly across these complex business environments.

Open‑Source and Commercial Parallel Strategy

The dual‑track approach is described as smart: the 35B version is fully open‑source to attract the developer community, while the 122B version remains closed‑source for enterprise customers, balancing ecosystem growth with commercial value.

Related Links

Official blog: https://hcompany.ai/holo3

Hugging Face model page: https://huggingface.co/Hcompany/Holo3-35B-A3B

API documentation: https://hcompany.ai/holo-models-api

GitHub Cookbook: https://github.com/hcompai/hai-cookbook/tree/main/holo3

The models are already accessible through H Company’s inference API, with documentation and quick‑start guides available online, providing enterprises with a highly cost‑effective solution for high‑performance computer‑use AI.