Hcompany Releases Holo4: Open-Weight Computer-Use Agent for Enterprise
On September 28, 2026, French AI startup Hcompany released Holo4, a family of open-weight models built specifically for computer use across every interface an enterprise workflow might require: graphical UIs, code execution, MCP tool calls, and direct business API access, all within a single model called the same way regardless of environment.
That “one model, all interfaces” design is the headline. Most computer-use models available today are trained for one interaction layer. GUI-focused models are blind without a screen. Tool-calling models get stuck when a task requires clicking through an interface that has no API. Real enterprise workflows combine both, often mid-task, and that mismatch is where most agentic deployments break down.
Holo4 was built to remove that split.
What Holo4 Is
Holo4 ships in two sizes. Holo4-27B is a dense model built on Qwen3.8-27B, targeting accuracy across long, multi-step workflows on web, desktop, and mobile. Holo4-35B-A3B uses a Mixture-of-Experts architecture based on Qwen3.6, trading some raw capability for better cost-per-token on high-throughput deployments. Both share identical architecture for tool access and are called identically by the hai-agents harness.
Hcompany also released Holotron4 Nano, applying the same post-training stack to NVIDIA’s Nemotron 3 Nano Omni model as part of its membership in the NVIDIA Nemotron Coalition. It targets smaller-footprint deployments while preserving generalist agentic behavior.
The core capability across all three: the model can click and type on a screen, write and execute code, call MCP tools, and hit business APIs. It uses whichever interface fits the step it is on, switching mid-task without the operator needing to route it to a different model.
Platforms confirmed on launch: desktops (macOS, Windows, Linux), the web, Android, code sandboxes, MCP servers, and business APIs.
Benchmark Scores and What They Actually Mean
Hcompany published benchmark results against OSWorld and OSWorld 2.0, the primary evaluations for long-horizon computer-use tasks.
On OSWorld 2.0, Holo4-27B scores 61.7%, compared to 81.8% for Claude Opus 5.5, according to the official Hcompany blog. The gap is meaningful, but the context matters: Holo4-27B is a 27-billion-parameter open-weight model running at a fraction of the compute cost of a frontier closed model. On the separate OSWorld benchmark (not version 2.0), Holo4-27B scores 85.2% at $0.08 per task, according to the model card on Hugging Face.
Hcompany also evaluates models on Agentic Task Factory, an internal held-out set of real business workflows across web, desktop, and MCP tools. These scores are not published against competitors but serve as Hcompany’s primary validation that the model handles production scenarios, not just academic benchmarks.
| Model | OSWorld 2.0 | OSWorld | Params | License |
|---|---|---|---|---|
| Holo4-27B | 61.7% | 85.2% | 27B dense | CC BY-NC 4.0 |
| Holo4-35B-A3B | 30.9% | N/A | 35B-A3B MoE | CC BY-NC 4.0 |
| Claude Opus 5.5 | 81.8% | N/A | Closed | API only |
For enterprise teams, the OSWorld 2.0 gap between Holo4 and frontier closed models is real. The more useful framing is whether Holo4 is accurate enough for the specific workflow at hand, and whether the cost and deployment flexibility advantages offset the accuracy gap on tasks where a 20-point difference materially matters.
How Hcompany Trained It
The training stack is worth understanding, because it explains both the generalist behavior and why Hcompany can compete with companies that have far more compute.
Holo4 was trained on 127 billion tokens, including successful agentic trajectories from desktop, web, MCP, API, and mobile environments. The source of those trajectories is the Agentic Task Factory, an internal pipeline that builds interactive environments and verifiable tasks from documentation alone, using screenshots of real websites and open-source software. At the time of launch, the factory had produced approximately 10,000 tasks across web apps, MCP servers, desktop environments, and hybrid GUI/MCP environments where the same state is reachable through both a screen and a tool call.
Training combined supervised learning on those trajectories with reinforcement learning on the verifiable tasks. The RL component is what builds robustness: the model learns to recover from failed actions rather than only imitating successful ones.
One architectural detail relevant to enterprise deployments: Holo4 supports reliable memory across hundreds of steps and can access a shell directly on the desktop machine. This changes the ceiling for what a single agent run can accomplish, because it removes the context-length cutoffs that truncate most current agentic systems mid-workflow.
Context window is 262,144 tokens across both sizes.
All training trajectories are open-sourced on Hugging Face and viewable step-by-step at trajectories.hcompany.ai, which is a transparency move with no peer among other computer-use model releases.
Why This Matters for Enterprise AI Teams
The computer-use category has been a growing focus for enterprise automation teams through 2026. Anthropic’s computer use Skills API reaching general availability in August established enterprise compliance infrastructure for the capability. Hark launched its handoff-focused computer-use agent for teams that want human-in-the-loop oversight on sensitive steps. Holo4 is the first open-weight model in this cohort that takes cross-interface generalization seriously from architecture rather than as an add-on.
For enterprise teams, the open-weight availability changes the deployment calculus in three specific ways.
Cost control at scale. At $0.08 per OSWorld task for the 27B model, Holo4 sets a concrete benchmark for what computer-use automation should cost at production volume. Teams running thousands of automation tasks per day can now model total cost of ownership against a published number, rather than extrapolating from API pricing at low volumes.
Self-hosted deployment. The CC BY-NC 4.0 weights on Hugging Face in FP8 and GGUF formats can run on enterprise infrastructure. For teams in regulated industries that cannot route screenshot data through a third-party API, self-hosted computer use is no longer theoretical.
Audit and reproducibility. The full trajectory dataset means teams building on Holo4 can inspect the training data for the behaviors they are deploying. Most AI model deployments in regulated contexts have no visibility into the training distribution. Hcompany’s open trajectories change that for computer-use workflows.
The gap with frontier closed models on OSWorld 2.0 is real, and teams choosing between Holo4 and Claude Opus 5.5 or GPT-6 Astra for their highest-stakes automation workflows should evaluate against their specific task distribution. But for the category of workflows where 60-85% accuracy is operationally sufficient, and where cost and deployment flexibility matter, Holo4 changes the competitive map.
For enterprise teams building toward AI-native operations, the arrival of a credible open-weight computer-use model shifts the infrastructure decision from “which API vendor” to “what accuracy do I actually need and what can I afford to run myself.” That is a better problem to have than the one that existed six months ago.
Commercial access is available now through the H Models API. The weights are live on Hugging Face. Reach out to talk through how computer-use automation fits your enterprise stack.