What Are Open-Weights Models? Open Weights vs Open Source, Explained

Open-weights models are AI models whose trained parameters — the "weights" — are published for download, so you can run the model on your own infrastructure instead of calling someone else's API. That single fact changes a lot: where your data goes, what you pay, and how much you can customize. But "open weights" is often confused with "open source," and the two are not the same thing. Getting the distinction right matters before you commit a budget or a roadmap to one.

This guide explains open-weights models in business terms. We cover a crisp definition, the real difference between open weights vs open source vs proprietary models, what you actually receive when you download one, why it matters for privacy and cost, and the misconceptions that trip up first-time buyers. Where licensing gets specific, we point you to official sources rather than guessing — because the details vary model by model.

What are open-weights models? A plain definition

An open-weights model is one where the developer releases the trained model's weights — the billions of numbers the model learned during training — as a downloadable file. With those weights and a compatible runtime, you can run the model yourself: on a cloud server you control, on-premises hardware, or in some cases a laptop.

The key word is "weights." You get the finished, trained brain of the model. What you typically do not get is the full recipe used to build it — the training datasets and the original training code — and the license attached to the weights may place limits on how you use them. That is the line that separates open weights from open source, and it is the single most important thing to understand on this page.

Open weights vs open source vs proprietary

Most AI models fall into one of three buckets. The terms get used loosely in marketing, so here is the practical distinction that matters for a buyer.

The Open Source Initiative (OSI) — the long-standing steward of the "open source" label — published its Open Source AI Definition in 2024 and has been explicit that releasing weights alone does not make a model open source. Open source, in the OSI sense, requires enough information to study, modify, and rebuild the system, including the training code and meaningful detail about the data. Open weights clears a lower bar.

• Proprietary / closed: You access the model only through an API or app. You never see the weights, cannot run it offline, and your data passes through the provider. Examples in spirit: the flagship commercial assistants offered purely as a service.
• Open weights: You can download and run the weights, fine-tune them, and self-host. But you may not get the training data or training code, and the license can restrict commercial use, redistribution, or specific use cases. Most "open" models businesses talk about live here.
• Open source (OSI-style): You get the weights plus enough of the training code and data information to genuinely study and reproduce the model, under a license that grants broad freedom to use, modify, and redistribute. This is the strictest and least common tier.

• Open Weights: not quite what you've been told — Open Source Initiative
• The Open Source AI Definition — Open Source Initiative

What you actually get when you download an open-weights model

"Downloading a model" sounds abstract until you see the package. Open-weights releases — usually hosted on a model hub — bundle a few distinct things, and each one matters for a different reason.

• The weights themselves: large binary files (often several gigabytes to hundreds of gigabytes) containing the trained parameters. This is the model.
• A model card: documentation describing what the model is good at, how it was evaluated, known limitations, intended uses, and — importantly — the license. Always read this before deploying.
• A license: the legal terms governing use. This ranges from permissive open-source licenses to custom "community" or "acceptable use" licenses with conditions. The license, not the marketing, defines what you may do.
• Config and tokenizer files: the supporting pieces a runtime needs to load and run the model correctly.

Why open-weights models matter for your business

For a small or mid-size company, the appeal of open-weights models is rarely about ideology. It is about four concrete levers: control, privacy, cost, and independence from a single vendor.

Because you run the model yourself, you decide where it lives and what touches your data. That is a meaningfully different risk posture than sending every prompt to a third-party API.

• Privacy and data control: Self-hosting an open-weights model means sensitive prompts and documents can stay inside your network. For regulated industries or confidential workflows, this is often the deciding factor.
• Cost predictability: API pricing scales with usage and can surprise you at volume. Running your own model converts that into largely fixed infrastructure cost, which can win at scale — though you take on the operational work. (See our cost guide for the real math.)
• Customization: You can fine-tune open weights on your own data to specialize the model for your domain, tone, or tasks — something closed APIs limit or don't allow.
• No vendor lock-in: The weights are yours to keep. If a provider changes pricing, deprecates a model, or shuts down, your deployment keeps working. That continuity is hard to buy from a closed API.

Common misconceptions about open-weights models

The phrase "open weights" invites a few assumptions that can get a business into trouble. Clearing them up early saves legal and engineering headaches later.

• "Open weights means free of all restrictions." Not necessarily. Many open-weights licenses include conditions — commercial-use thresholds, restricted use cases, or attribution and redistribution rules. Open to download is not the same as free to do anything.
• "Open weights means open source." No. As the OSI makes clear, open weights typically omit the training data and code that open source requires. Treat them as related but distinct tiers.
• "Self-hosting is automatically cheaper." It can be, but you trade API fees for hardware, ops, and maintenance. The break-even depends on your volume and your team.
• "All open-weights models are equally capable." Capability varies widely across families and sizes. Pick based on benchmarks and your actual task, not on the label alone.

Well-known open-weights model families

Several major model families are commonly distributed as open weights, and they are a good starting point for evaluation. Capabilities, sizes, and — critically — license terms differ across and within each family, and they change over time. Treat the names below as a map, and confirm the specifics on each model's official model card before you build on it.

Notable open-weights families businesses evaluate today include Meta's Llama, Mistral, Alibaba's Qwen, DeepSeek, Google's Gemma, and Microsoft's Phi. Some are released under permissive licenses; others use custom community or acceptable-use licenses with conditions. The only reliable source for what you may do with a given model is its published license and model card — not a third-party summary, including this one.

Conclusion: open weights vs open source, and what to do next

Open-weights models give businesses something proprietary APIs cannot: the ability to download, run, customize, and keep an AI model on their own terms. That unlocks real advantages in privacy, cost control, customization, and freedom from vendor lock-in. The catch is precision — open weights vs open source is not a cosmetic distinction. Open weights means you can run the model; open source (in the OSI sense) means you also get the code and data to fully study and rebuild it, under a permissive license. And neither label guarantees you may do whatever you want — the model card and license decide that.

If open-weights models look like a fit for your roadmap, the next steps are to shortlist a few families, read their licenses carefully, and run the cost and safety math for your situation. That is exactly the kind of evaluation we help businesses get right.