Key Takeaways
- "Open weight" means the trained model file is public and runnable by anyone β it says nothing about whether the license meets the Open Source Initiative's (OSI) definition.
- "Open source," in the strict sense, requires an OSI-approved license β such as Apache 2.0 or MIT β with no restriction on commercial use, business scale, or field of application written into the license itself.
- Llama (Meta) and Gemma (Google) are open weight but not OSI-approved open source; both licenses attach a usage-scale or conduct restriction to the weights.
- Most current DeepSeek, Phi, GLM, and gpt-oss releases ship under a genuine OSI-approved license (MIT or Apache 2.0); Qwen and Mistral do too for most current sizes β always verify the specific checkpoint.
- A small number of projects, like OLMo, publish the training code and data alongside the weights, not just the weights themselves β the rarer case sometimes called "fully open."
- For personal, local use, the distinction rarely changes anything in practice. For a commercial product, a redistributed fine-tune, or a SaaS deployment, it can determine whether you need a separate license.
π In One Sentence
"Open weight" means a model's trained parameters are downloadable and runnable; "open source" in the strict sense additionally requires an OSI-approved license with no commercial-use, scale, or field-of-use restriction.
π¬ In Plain Terms
Downloadable does not automatically mean open source. Check the actual license file β not the marketing page β before assuming a model is free to use commercially, redistribute, or rename.
"Open Weight" Means Downloadable Parameters, Not Open Permissions
"Open weight" describes a model whose trained parameters β the numeric weights that make the model function β are published for anyone to download and run. That is the entire claim. It does not say anything about the training code used to build the model, the dataset it was trained on, the exact training recipe, or whether the accompanying license meets any external standard for openness.
In practice, most open-weight releases ship with a custom license written by the vendor rather than a standard open-source license. Meta's Llama Community License and Google's Gemma Terms of Use are the two most common examples on this site's model comparisons: both let you download and run the weights, both permit commercial use for most organizations, and both attach a condition that a standard open-source license would not β Llama a usage-scale threshold (a monthly-active-user cap, above which a separate agreement is required), and Gemma a conduct-based Prohibited Use Policy layered on top of its own Terms of Use.
None of this makes "open weight" a bad or deceptive label β it is an accurate description of what actually gets released. The problem is only when "open weight" gets shortened to "open source" in casual writing, because the two labels answer different questions: open weight answers "can I download and run this file?" and open source (in the strict sense below) answers "does the license itself meet an external openness standard with no usage-scale or field-of-use catch?"
"Open Source" Requires an OSI-Approved License, Not Just Downloadable Weights
A model is "open source" in the strict sense only if its license is approved by the Open Source Initiative (OSI) β the organization that maintains the Open Source Definition. That definition specifically prohibits licenses that discriminate against fields of endeavor (for example, "free for research, not for commercial use") or against groups of people, and it prohibits restrictions tied to the scale of the licensee's business. A usage-based cap like a monthly-active-user threshold fails that test even when the weights themselves are freely downloadable β which is exactly why Llama and Gemma, described above, do not qualify as OSI-approved open source despite being widely downloadable.
Two licenses cover the large majority of genuinely open-source local LLM releases: the Apache License 2.0 and the MIT License, both OSI-approved, both permitting commercial use, modification, and redistribution with only a standard notice-and-attribution requirement, and neither carrying a usage-scale or revenue threshold. When a model family's license page names one of these two licenses for the specific checkpoint you are looking at, that checkpoint is open source in the strict sense described here.
One caveat worth knowing: the Open Source Initiative has also published a separate Open Source AI Definition, which goes further than a permissive code license and asks for meaningful transparency about training data and the training process itself β not just a license text on the weights. Almost no model on this comparison meets that fuller bar; OLMo, covered below, comes closest among the families listed here. That distinction is beyond what this article covers in depth, but it is worth knowing if a vendor markets a model as "open source AI" and you want to check what that claim is actually resting on.
A Few Releases Go Further and Publish the Training Data Too
Weights-only openness and fully open releases are not the same thing, and the gap between them is usually the training data and code, not the license text. Most of the models in the comparison table below are open source in the license sense β the weights ship under Apache 2.0 or MIT β but the dataset used to train them, the exact data mixture, and the training pipeline itself stay proprietary. OLMo, from the Allen Institute for AI, is the clearest example on this list of a model that goes further: it publishes the training code and a substantial share of the training data alongside the weights, under Apache 2.0. That additional layer of openness is genuinely rare in the current local LLM landscape β treat any "fully open" claim about a model family as unverified until you can point to the actual published dataset and training code, not just a license badge on the model card.
How Local LLM Licenses Compare: Open Weight vs. Open Source vs. Fully Open
No two of these license structures are identical, and several families ship different licenses on different checkpoints β the license is set per model file, not per brand. Use this table as an orientation, then check the LICENSE file for the exact checkpoint before a commercial deployment.
Model family | Vendor | License | Category |
|---|---|---|---|
| Llama 3 / 4 | Meta | Llama Community License | Open weight only (not OSI-approved) |
| Qwen | Alibaba | Apache 2.0 (most) / custom (some) | Mostly open source β check per checkpoint |
| Mistral / Mixtral | Mistral AI | Apache 2.0 (most) / Modified MIT (some) | Mostly open source β check per checkpoint |
| Gemma | Gemma Terms of Use | Open weight only (not OSI-approved) | |
| DeepSeek | DeepSeek | MIT (current flagship) | Open source license (weights only) |
| GLM | Zhipu / Z.ai | MIT / Apache 2.0 (most current) | Open source license (weights only) |
| gpt-oss | OpenAI | Apache 2.0 | Open source license (weights only) |
| Kimi K2 | Moonshot AI | Modified MIT | Open-source-style, one added clause |
| Phi | Microsoft | MIT | Open source license (weights only) |
| OLMo | Ai2 | Apache 2.0 | Fully open β weights, code, and data |
This table describes the general license structure as commonly published. It is not legal advice, and it is not exhaustive: several families here (Qwen, Mistral, and to a lesser extent GLM) have released specific checkpoints under more restrictive terms than the family's usual license, and license terms change between model versions. Kimi K2's Modified MIT license adds a clause to plain MIT β read it, rather than assuming "Modified MIT" behaves identically to MIT. Always confirm the LICENSE file attached to the exact checkpoint you plan to deploy. For a deeper, deployment-scenario-by-scenario breakdown of commercial-use terms across these and other families, see the Local LLM Commercial License Guide.
The License Only Matters Once You Go Beyond Personal Use
If you are running a model locally for your own use β chatting, coding assistance, writing, experimenting β the open-weight-vs-open-source distinction almost never changes what you can actually do. Every model in the comparison table above permits personal, local use under its current license. Download it, run it through Ollama or llama.cpp, fine-tune it on your own machine for your own purposes, and the practical difference between "open weight" and "open source" is close to zero.
The distinction starts to matter the moment you build something for other people. If you are shipping a commercial product, offering a hosted API or SaaS built on the model, redistributing the weights (including a fine-tuned derivative), or approaching a scale threshold a license defines (like Llama's monthly-active-user cap or Mistral's revenue-cap checkpoints), the license category tells you what to check next: an OSI-approved license like Apache 2.0 or MIT generally has no such catch; an open-weight license like Llama's or Gemma's might.
This is not legal advice, and none of the license summaries here substitute for reading the actual license text attached to the specific checkpoint you plan to deploy. For any commercial decision, verify the current license and consult counsel before shipping.
Frequently Asked Questions
Is Llama open source?
Not in the strict, OSI-approved sense. Llama is open weight: the parameters are freely downloadable and commercial use is permitted for most organizations, but Meta's Llama Community License adds a usage-scale threshold (a monthly-active-user cap) that the Open Source Initiative's definition does not allow in an approved open-source license.
What does "open weight" mean?
It means a model's trained parameters are published for anyone to download and run. It does not say anything about whether the training code or training data are published, or whether the license meets any external open-source standard β check the specific license text for those answers.
Can I use DeepSeek commercially?
DeepSeek's current flagship releases ship under a plain MIT license, which permits commercial use, modification, and redistribution with no usage threshold. Some earlier DeepSeek checkpoints shipped under a different, more restrictive license, so always check the LICENSE file for the exact checkpoint before deploying it commercially β this is not legal advice.
Which local LLMs are truly open source?
Most current DeepSeek, Phi, GLM, and gpt-oss releases ship under a genuine OSI-approved license (MIT or Apache 2.0), and most current Qwen and Mistral sizes do too. OLMo goes a step further and publishes the training code and data alongside the weights, not just the weights themselves β the rarer, fully open case.
Why isn't Gemma considered open source?
Gemma's weights are freely downloadable and commercial use is generally permitted, but Google distributes it under its own Gemma Terms of Use plus a separate Prohibited Use Policy rather than a standard OSI-approved license. That conduct-based restriction is exactly the kind of clause the Open Source Definition prohibits, which puts Gemma in the open-weight category rather than open source.
Does it matter which license a model uses if I only run it locally for myself?
Almost never. Every model family covered in this article permits personal, local use under its current license. The license category becomes relevant once you build a commercial product, redistribute the weights, or approach a usage or revenue threshold a specific license defines.