Holo4 is here. On September 28, 2026, Paris-based AI company H Company released Holo4, its new generation of computer-use agent models: a 27B dense version and a 35B-A3B mixture-of-experts version, with weights published on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF precisions and API access opened in parallel. The company also released Holotron4 Nano, an updated version of Holotron 3.
Holo4's pitch fits in one sentence: it is not picky about interfaces. Clicking and typing on a screen are table stakes; it also writes and runs its own code and calls MCP or API tools, choosing whichever approach fits the task at hand. More importantly, desktops, the web, Android, code sandboxes, and business APIs all run on the same model with the same calling convention — no swapping models per platform.
In its official blog post, H Company says Holo4 was trained with supervised learning plus reinforcement learning on a large set of real and synthetic environment tasks, some generated by its own Agentic Task Factory. The goal is explicit: built for real enterprise business workflows, not for benchmark leaderboards.
Benchmarks and cost: cheap, but don't stop at the score
The official numbers: Holo4-27B reaches 85.2% on OSWorld at about $0.08 per task, and 61.7% on the harder OSWorld 2.0. H Company itself concedes this trails top closed models like Opus 5.5 (81.8%) by a clear margin — but the parameter counts and costs differ by more than an order of magnitude.
One detail deserves its own mention: H Company has open-sourced the complete trajectories behind all its public benchmark scores, with every step replayable. At a time when benchmark gaming is a constant controversy, that at least makes the scores verifiable.
On API pricing, the 35B-A3B costs $0.30 per million input tokens and $2.00 per million output tokens; the 27B is labeled research-only at $0.40 per million input tokens. This price band is clearly aimed at running agents cheaply.
Read the license first: 27B is non-commercial, 35B is commercial
Anyone planning to self-host the weights should read the license before doing anything else. Holo4-27B is released under CC BY-NC 4.0 — non-commercial use only. The 35B-A3B is Apache 2.0 and can be used commercially. In other words, the 27B fits research, evaluation, and internal experiments; for production, it's the API or the 35B-A3B. The licensing split directly affects model choice — don't discover it halfway through deployment.
In context: open weights are the differentiator
The computer-use agent space is no longer short of players — OpenAI, Anthropic, and Google all have their own offerings. Holo4's edge is not its absolute score but three things: open weights, a unified all-interface approach, and open-sourced evaluation trajectories. For enterprise users, that means self-hosting, auditable model behavior, and reproducible failures — precisely what closed offerings struggle to provide.
The weaknesses are real, too: absolute capability still trails top closed models by a generation, performance on Chinese-language software environments remains unverified, and the 27B's non-commercial license limits direct deployment. Holo4 reads more like a signal: competition in computer-use agents is shifting from 'who scores highest' to 'who is more usable, cheaper, and more controllable.' This site previously reported that open-weight model mentions in enterprise earnings rose 6x year over year — Holo4 lands squarely on that trend. Enterprises don't want the smartest model; they want one whose math works out.
