Open-weight model

What is an open-weight model?

A model's weights are the numbers it learned in training. Publish them and anyone can download the model, run it on their own servers, inspect it and fine-tune it. Stanford HAI puts it plainly: an open-weight model "is an AI model whose core components are publicly released, allowing anyone to download it."

The opposite is a closed model, such as Claude or GPT, which you can only use through its maker's API. You send your data to the model; you never hold the model.

How open-weight models work

An open-weight release usually comes as a set of files on a model hub such as Hugging Face, with a license that says what you may do with them. To use one, you load the weights into serving software such as vLLM on hardware you control, or fine-tune them on your own data first.

Open-weight vs open source vs closed models

The Open Source Initiative draws the line: open weights "refer to the final weights and biases of a trained neural network," and differ from open source AI because they don't include the training code or the training dataset (or, where the data can't be shared, full details of it). In everyday use, "open source LLM" usually means an open-weight model.

A short history of open-weight models

Sources: Meta, Mistral AI, H2O Danube report, and the DeepSeek-R1 and gpt-oss model cards. Open-weight models now come from labs in the US, Europe and Asia, and open models rival frontier models on accuracy for many tasks, so owning a model no longer means settling for less.

Open-weight or closed: how to decide

Choose an open-weight model when:

• The data is sensitive or regulated and shouldn't go to a model provider.
• You need to run on-premises, in-region or air-gapped.
• You run the same task at high volume, where per-token bills add up.
• You need to fine-tune the model, freeze a version, or prove to a regulator exactly which model made a decision.

A closed model can make sense for early experiments, for open-ended tasks where only the largest models do well, or when you have no team to run models yourself. Many organizations use both, with a router sending each job to the right one.

Open-weight licenses

Open-weight doesn't mean free for any use. The license decides. Permissive licenses such as Apache 2.0 allow commercial use, which is how H2O Danube, H2O Mississippi and H2O Lightning are released. Others restrict it: TimeH2O-v1's open weights are released under the H2O Model Non-Commercial License for research and evaluation, and production use needs a license from H2O.ai. Read the license before you build on a model.

Why organizations use open-weight models

• Ownership. A model you hold can't be retired, repriced or changed under you, the core of sovereign AI.
• Privacy. Your data never goes to a model provider, and the model can run on-premises or air-gapped.
• Customization. You can fine-tune the weights on your own data, often turning a small language model into an expert at one job.
• Cost. You pay for hardware you control rather than per token.

How to evaluate an open-weight model

• License. What it allows: commercial use, redistribution, fine-tuned versions, and any limits on users.
• Fit. Benchmarks close to your task first, then a test on your own data, which matters more than any leaderboard.
• Provenance. The model card: who trained it, on what data, and with what known limits.
• Safety. Red-team it before production. Open toolkits such as Project Moonshot, from the AI Verify Foundation, test and red-team language models.
• Support. Who patches it, how you'll update it, and who you call when it misbehaves.

Open-weight models in regulated industries

Banks, insurers, health systems and government agencies favor open-weight models for a simple reason: their model risk teams can validate a model they hold, test it on their own data, and freeze the version they approved. A rented model can change under them. H2O.ai built an industry-first generative AI model risk management framework and trained risk teams at 14 banks, including Commonwealth Bank of Australia, Wells Fargo, Barclays and HSBC.

Risks and limits of open-weight models

• Security is yours. Patching, access control and monitoring move from the provider to you.
• Misuse. Anyone can download the same weights. The US NTIA's 2024 report found the evidence did not yet justify restricting open model weights, and recommended watching for emerging risks.
• The biggest models stay closed. The largest frontier models are generally offered only through an API.
• License terms vary. Some limit commercial use, user numbers or what you build with the model.

Common myths about open-weight models

• "Open-weight means free to use." The license decides, and some forbid commercial use.
• "Open-weight means open source." Usually the training data and code stay private.
• "Open-weight models are always weaker." Open models now rival frontier models on accuracy for many tasks, and a fine-tuned open model often beats a general closed one on a narrow job.
• "You need a huge cluster." A small language model with open weights runs on a single GPU server or a laptop.

Open-weight model examples

• Meta Llama and Mistral, general-purpose language models in many sizes.
• Alibaba Qwen and DeepSeek, open-weight models from China used widely around the world.
• H2O Danube, open-weight small language models under Apache 2.0.
• H2O Mississippi, an open-weight vision-language model with 2.4 million monthly downloads (July 2026).
• H2O Lightning, open-weight decision models; H2O-Lightning-4B ranked first among 104 open-weight systems on JevBench as of October 7, 2026.

Where to get open-weight models

• Hugging Face, the largest public hub, where each model has a model card and a license.
• The makers' own sites and GitHub, with release notes and technical reports.
• AI platforms that package and run them, so you don't wire up serving, security and monitoring yourself.

Open-weight models in H2O AI

H2O.ai publishes its own open-weight models on Hugging Face and documents their methods in three technical reports on arXiv. H2O AI Cloud runs any open-weight model in your own environment, including on H2O AI Hybrid Cloud. H2O Cascade fine-tunes and distills them on your data, and the H2O AI Super Agent routes each job to the right model. Results for H2O.ai's open-weight models, including H2O Lightning on JevBench and TimeH2O on fev-bench, are on the H2O.ai benchmarks page.