Is Holo4 Free to Use? License and Commercial Use Explained
Holo4's open weights from H Company are free to download, but licensing and API access affect what you can actually ship with them.
What is Holo4?
Holo4 is a family of open-weight models from the Paris-based lab H Company, built to act as computer-use agents. Instead of specializing in just clicking through graphical interfaces or just calling APIs and tools, Holo4 is trained to work across all three “lanes” of agent action: GUI interaction (looking at screenshots and clicking or typing), writing and running code, and calling tools through MCP servers or APIs. The lineup includes a dense Holo4 27B model built on Qwen3.8, a 35B mixture-of-experts model (3B active) built on Qwen3.6, and a smaller variant called Holotron Nano built on Nvidia’s Nemotron 3 Nano Omni recipe. Weights for all of them are published on Hugging Face in multiple quantizations, and H Company also offers a hosted API for running the models without local hardware.
The question that matters for anyone planning to build on Holo4 isn’t whether the weights are available. They are. It’s what license governs them, and whether that license allows commercial deployment or only research and experimentation.
TL;DR
- Holo4’s weights are openly downloadable on Hugging Face in several sizes and quantization formats, including a 27B dense model and a 35B mixture-of-experts variant with 3B active parameters.
- Open weights do not automatically mean open commercial rights, so anyone planning to ship a product needs to check the specific license terms attached to the model files before deploying.
- H Company runs its own hosted API alongside the open weights, which is the path most likely intended for commercial or production use rather than self-hosting under a restrictive license.
- Local inference is viable but resource-heavy, since the full bfloat16 weights run to tens of gigabytes and screenshot-based GUI tasks consume KV cache fast, meaning serious VRAM headroom is needed.
- The dense 27B model consistently outperforms the mixture-of-experts 35B model on agentic benchmarks like OSWorld, which matters when deciding what to self-host versus what to run through the API.
- Research and prototyping use cases are clearly supported by the open release, but teams building commercial products should treat the hosted API as the lower-risk option until license terms are confirmed for their use case.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Why does the license matter for an agentic model like this?
Agentic models that click, run code, and call tools sit closer to “software that takes actions” than a typical text generator. That raises the stakes on licensing in two ways. First, if a license restricts commercial use, deploying the model inside a paid product, a customer-facing agent, or any revenue-generating workflow could violate those terms even if the weights are sitting right there on Hugging Face. Second, because Holo4 is designed to automate real tasks (filling out forms, extracting data from legacy software, posting to Slack, running scripts against live systems) it’s the kind of model teams want to wire into actual business processes almost immediately. That’s exactly the scenario where license terms stop being a formality and start being a legal question worth checking before go-live.
Open weight releases from labs building commercial products often follow a familiar pattern: free access for research and personal projects, with a separate commercial license or a paid API required for production use. This lets the lab build goodwill and adoption in the research and developer community while still monetizing companies that want to deploy at scale.
What’s actually free to do with Holo4?
Based on how H Company has structured the release, with open weights on Hugging Face alongside a parallel hosted API, downloading the model and running it locally for testing, research, benchmarking, or internal prototyping is the use case the open release is clearly built for. You can pull the 27B dense model or the smaller mixture-of-experts version, run it on your own hardware, and experiment with building an agent harness around it, as demonstrated with GUI-only tasks, tool-calling through MCP, and code execution.
What’s less clear without reading the exact license text attached to the weights is whether that same self-hosted deployment is permitted once it’s generating revenue, whether through an internal tool that saves a company money, a SaaS product, or a client-facing agent. Non-commercial research licenses are common for labs that also sell API access, and the structure here (open weights plus a dedicated paid API) is consistent with that model, though the brief doesn’t give exact license terms to quote. Anyone evaluating Holo4 for a commercial product should read the license file shipped with the specific model checkpoint they intend to use, rather than assuming “open weights” means “commercial use permitted.”
How does the paid API fit in?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
H Company offers its own API for Holo4, which runs alongside their hosted agent harness and SDK called High Agents. That harness is built to plug into other tools, including integration with Claude Code over MCP, letting a larger orchestration model hand off screen-heavy or tool-calling work to Holo4 while it handles higher-level planning. For teams that don’t want to manage GPU infrastructure, or that want clear commercial terms without parsing a weights license, the API is the straightforward path. It also sidesteps the practical headaches of self-hosting: screenshots fed into a vision-capable agent consume context and KV cache quickly, so running the 27B model locally at scale requires real VRAM headroom, not just enough to load the weights.
Is it worth running Holo4 locally versus using the API?
For research, benchmarking, or building a prototype where cost control and experimentation matter more than production uptime, running Holo4 locally makes sense, provided the hardware is there. The full bfloat16 version of the larger model is a multi-ten-gigabyte download, and GUI-based agent loops that pass screenshots back into the model every step will eat through context fast, so a workstation-class GPU with substantial memory is the realistic minimum for smooth operation.
For commercial deployment, the calculus shifts. Benchmarks comparing Holo4 to frontier closed models on tasks like OSWorld show it lands meaningfully behind the top proprietary agents on raw accuracy, but at a fraction of the inference cost, since running frontier vision models on every screenshot in an agent loop gets expensive quickly. That cost advantage is part of Holo4’s pitch. But realizing it at the lowest commercial risk likely means going through H Company’s official API rather than self-hosting under a license that may not cover revenue-generating use. Until the exact commercial terms for self-hosted deployment are confirmed, treating the API as the default production path and the open weights as the research and prototyping path is the safer assumption.
Frequently Asked Questions
Are Holo4’s weights completely open source?
The weights are published openly on Hugging Face with multiple quantization options, which makes them freely downloadable and runnable. Whether “open” extends to unrestricted commercial use depends on the specific license attached to the release, which should be checked directly rather than assumed.
Can I use Holo4 in a commercial product for free?
That depends on the license terms tied to the specific model checkpoint. Labs that release open weights alongside a paid hosted API commonly restrict free use to research or non-commercial purposes, reserving commercial deployment for the paid API or a separate commercial license. Confirm the terms before shipping a product built on self-hosted Holo4 weights.
What’s the difference between the Holo4 27B and 35B models?
The 27B is a dense model built on Qwen3.8, while the 35B is a mixture-of-experts model with 3 billion active parameters built on Qwen3.6. On agentic benchmarks involving screenshots and GUI tasks, the dense 27B model performs substantially better than the mixture-of-experts variant, suggesting the larger active parameter count matters more for visual, image-heavy reasoning.
Do I need the API, or can I self-host everything?
Self-hosting is possible if you have adequate GPU hardware, since the weights are downloadable and the model runs through standard local inference setups. The practical constraints are VRAM and KV cache headroom for screenshot-heavy tasks, plus whatever license restrictions apply to commercial self-hosted use. The API removes both the hardware burden and the licensing ambiguity for production deployments.
How is Holo4 different from earlier H Company models?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Earlier Holo models focused on function calling. Holo4 is trained to treat GUI interaction, code execution, and tool calling as equally first-class actions, and to choose between them dynamically based on which approach solves a given step most efficiently, rather than relying on one mode with the others bolted on.
