French AI company H Company released the Holo4 model family on Hugging Face on September 28, built specifically as "computer-use agents": they read screenshots, click the mouse, type on the keyboard, and can also call code, MCP and APIs to complete tasks. This generation ships two flagship versions — the 27-billion-parameter dense model Holo4-27B, and the mixture-of-experts model Holo4-35B-A3B, with 35 billion total parameters and roughly 3 billion activated per pass. Both sets of weights are released under the Apache 2.0 license.
There's also a Holotron4 Nano, trained on top of Nvidia's Nemotron 3 Nano Omni, sized at 30B-A3B.
H Company is based in Paris, founded in 2024 by a team that includes several former DeepMind researchers. It raised roughly $220 million in seed funding the same year it launched. The Holo series is its flagship product line, and the company's blog says Holo4 was trained onward from its previous generation.
Benchmarks and positioning
H Company leans mainly on OSWorld 2.0, a benchmark that has models complete multi-step tasks inside a real desktop environment. Its blog post gives this comparison:
- Holo4-27B: 61.7%
- Claude Opus 5.5: 81.8%
- Claude Opus 5: 70.2%
- GPT-5.6 Sol: 66.2%
Judging by those numbers, Holo4-27B is now close to the previous generation's closed flagship, though it still trails the current strongest model, Opus 5.5, by a clear margin. H Company's framing is that Holo4 "lags only the strongest closed models" on long-horizon tasks, while using several orders of magnitude fewer parameters and costing far less per task.
The cost figure appears in the model card: Holo4-35B-A3B scores 34.5% on AutomationBench at roughly $0.02 per task. Costs for other models in the blog's charts are shown only as plotted points, without exact figures.
Demo scenarios include 3D modeling, game design, and enterprise business-process automation. All these scores were produced inside H Company's own testing framework; the company also notes that the models are meant to run with its own hai-agents framework, and no independent reproduction has surfaced yet.
What this means for developers in China
Start with the foundation. The model card for Holo4-35B-A3B lists Qwen3.6-35B-A3B, from Alibaba, as its base, with a maximum context of 262,144 tokens. Chinese teams are already familiar with deploying, quantizing and optimizing inference for this architecture, so swapping in Holo4's weights shouldn't require much change to existing tooling.
On whether it can run locally: Holo4-27B ships in BF16, FP8, NVFP4 and Q4 GGUF formats. Quantized to 4-bit, it could plausibly fit on a consumer GPU with 24GB of VRAM — a rough estimate that still depends on screenshot resolution and context length. The 35B-A3B variant activates only about 3 billion parameters per pass, which helps inference speed.
Computer-use agents run into an unavoidable problem: they need to keep sending screenshots to the model. With closed models like Opus or GPT, those screenshots have to go to overseas clouds, which many finance, government and manufacturing intranets simply don't allow. The Apache 2.0 license permits commercial use and modification, so an open model that can run in a local data center and still clears 60% gives those environments a realistic option.
H Company also offers a hosted service through its own H Models API; pricing hasn't been announced yet.
Sources: H Company's official blog, Hugging Face model card, CocoLoop; OSWorld 2.0 scores follow H Company's blog, AutomationBench results and per-task cost follow the model card — all figures are vendor-reported.