Octopi / Guides / Open source vs open weight
Guide
Open-source vs open-weight AI models: what’s the difference?
Open-weight means you can download the trained model — its weights — and run it on your own hardware. Open source, as the Open Source Initiative defines it for AI, also means the freedom to use, study, modify and share it, with enough information about its training data and code to do so. Many popular models are open-weight but not open source by that definition, so read the license.
What’s the difference between open-weight and open-source AI?
A trained AI model is, at its core, a large file of numbers called weights (or parameters). Anyone with the weights and a program to run them can use the model. What else is released, and under which license, separates the labels.
Used through a service
You send your input to the provider’s service and get an answer. You never get the weights; the provider decides which version you use.
Download it and run it
The weights are published for anyone to download. The license may still restrict how you use them, and the training data and code are often not released.
The full set of freedoms
Under the OSI’s definition: the freedom to use, study, modify and share the model, plus enough information about its data, code and weights to do so.
“Open source,” used loosely
In articles and on product pages, “open-source model” often just means open-weight. When it matters, check the license and what was released.
So every open source model is open-weight, but many open-weight models are not open source. The gap is about what the license lets you do and what you can learn about how the model was made.
What gets released with an AI model?
An AI release can include several parts. Here is what each label usually implies.
| Part | What it is | Open-weight | Open source AI (OSI) |
|---|---|---|---|
| Weights | The trained model itself | Downloadable | Included |
| Code to run it | Software that loads the weights and runs them | Often provided, or a common runner can load it | Included |
| Training code | The code used to train the model | Often not released | Included |
| Training data | What the model learned from | Often not released, or only described briefly | Enough information about the data |
| License | What you may do with all of the above | Anything from permissive to research-only | Must allow use, study, modification and sharing |
Releases differ from one publisher, and one model, to the next — hence “often.”
What is the Open Source AI Definition?
The Open Source Initiative (OSI) published the Open Source AI Definition 1.0 in October 2024.
In plain words, it asks that an AI system come with four freedoms: to use it, to study how it works, to modify it, and to share it. Because those freedoms mean little with only a file of numbers, it also asks for enough information about the training data, the code and the weights to make them practical.
So a model can be open-weight without meeting the definition — for example when its license restricts some uses, or its training data and code are neither released nor described. A permissive license on the weights, such as Apache 2.0, settles the license part, not the data and code parts.
Open-weight model licenses: can you use them commercially?
Most open-weight models use one of four kinds of license. One family can use different licenses for different sizes.
| License type | Examples | Commercial use | What to watch |
|---|---|---|---|
| Permissive (Apache 2.0, MIT) | Mistral 7B; Qwen2.5-7B, Qwen2.5-14B and Qwen3.5-9B (Apache 2.0) | Allowed | Keep the license and notices when you redistribute |
| Custom community license | Meta’s Llama 3.1 (Llama 3.1 Community License) | Allowed, with conditions | A separate license for products or services above 700 million monthly active users; an acceptable-use policy |
| Terms of use with prohibited uses | Google’s Gemma (Gemma Terms of Use) | Set by the terms — read them | A prohibited-use policy that applies to what you build |
| Research license | Qwen2.5-3B (Qwen Research License) | Not without a separate license: research or evaluation use only | For business use, ask the publisher for a commercial license or pick another model |
Check the exact model and size — licenses can differ within one family, and between versions. This is general information, not legal advice.
What should a business check before using an open-weight model?
Before a model goes into a product or a daily workflow, read its license with these questions.
Is it allowed at all?
Permissive licenses allow it, custom licenses may attach conditions, research licenses may not.
What must you pass on?
Many licenses ask you to keep the license text and notices when you redistribute the model or ship it in a product.
Is your use on a prohibited list?
Llama 3.1 has an acceptable-use policy and Gemma a prohibited-use policy. Read them against what you plan to do.
Does the license change at a size?
Llama 3.1’s license calls for a separate license above 700 million monthly active users. Look for clauses like this.
Can its answers train other models?
If you plan to train or improve another model on a model’s outputs, check whether the license addresses that.
Can you share what you build?
If you share or sell a fine-tuned version, check what the license says about derivatives, naming and passing its terms on.
Where do the weights come from?
Download from the publisher’s official page or your runner’s official library, and compare checksums where published.
Where does your data go?
The license doesn’t say. That depends on where the model runs and on the app built around it.
Why do open weights matter for privacy and sovereignty?
The license decides what you may do. Downloadable weights decide what you can do — that is where privacy and control come in.
Run it where your data is. You can run the model on your own computer or server, so the text you give it isn’t sent to a model provider to be processed. That is the core of running AI locally on a Mac.
Keep a copy. A hosted model can be changed or retired by its provider. Open weights on your own disk keep behaving the same way.
Choose and switch. You pick the model, the version and the machine, and can move to another model without asking anyone. Those are the controls behind sovereign AI: who holds the weights, who runs the compute, and whether you can switch.
Open weights are not a privacy guarantee on their own: the app around the model decides what else goes online, and running a model yourself doesn’t remove your duties under laws such as the GDPR. See AI, the GDPR and customer data and local AI vs cloud AI.
How do you check a model’s license in 2 minutes?
Name the exact model
Write down the family, version, size and variant — for example Qwen2.5-7B-Instruct, not just “Qwen.” Licenses can differ within one family.
Open the model card
On the publisher’s official page (often on Hugging Face), the model card names a license, and the files usually include a LICENSE. Read the file, not just the label.
Read the clauses that matter
Commercial use, any acceptable-use policy, user thresholds, and what you may do with copies and fine-tunes. If anything is unclear, ask a lawyer.
Already running the model? In Ollama, ollama show <model> --license prints the license that came with it, and a model’s page in the Ollama library usually lists it too. If the two disagree, go with the publisher’s model card.
Where Octopi fits
Octopi is one example of an app built on open-weight models: a free Mac app that turns a plain-English task into a finished spreadsheet, with a chat that runs on your Mac.
Open-weight models through Ollama
The model runs on your Mac through Ollama, the open-source engine that serves it, or on an Ollama server you run elsewhere. The chat refuses models Ollama runs in its own cloud.
Qwen2.5-7B-Instruct, Apache 2.0
The default for Macs with 12 GB of memory or more is published under the Apache 2.0 license, as are the Newest size, Qwen3.5-9B, and the Advanced size, Qwen2.5-14B.
Licenses vary — check each one
The Light size is a Qwen2.5-3B model, under the Qwen Research License, which allows research or evaluation use only. Llama 3.1, Mistral and Gemma 2 are offered too, plus any model already in your Ollama.
Free, but not open source
Octopi is a free app with no account needed; it is not open-source software. Its model is on your Mac, and it goes online for the web searches and CRM look-ups you ask for.
More on the chat: Chat with open-source AI models on your Mac.
Questions
What is the difference between open-source and open-weight AI?
An open-weight model’s weights can be downloaded and run on your own hardware, under the publisher’s license. Open source AI, as the Open Source Initiative defines it, also means the freedom to use, study, modify and share the model, with enough information about its data and code to do so. Every open source model is open-weight, not the reverse.
Is Llama open source?
Llama 3.1’s weights are downloadable, so it is open-weight. Its license, the Llama 3.1 Community License, permits commercial use with conditions — including a separate license above 700 million monthly active users and an acceptable-use policy — so open-weight is the more precise label than open source in the OSI sense.
Is Qwen open source?
It depends on the model. Qwen2.5-7B, Qwen2.5-14B and Qwen3.5-9B are published under Apache 2.0, a permissive license; some other Qwen2.5 sizes use Qwen’s own licenses — the 3B, for example, uses the Qwen Research License, which allows research or evaluation use only. Check the model card of the exact model and size.
Can I use open-weight models commercially?
Often, but it depends on the license of the exact model and size. Permissive licenses such as Apache 2.0 allow commercial use; custom licenses such as Llama 3.1’s allow it with conditions; research licenses may limit or rule it out. This is general information, not legal advice.
What is the Open Source AI Definition?
The Open Source Initiative’s definition of open source AI; version 1.0 came out in October 2024. It asks for the freedom to use, study, modify and share an AI system, with enough information about its training data, code and weights to make those freedoms practical.
Does running an open-weight model locally keep my data private?
For the model step, yes: a model running on your own computer doesn’t send your text to a model provider. The app around it may still go online, so check what it sends. Running locally also doesn’t remove your own data-protection duties.
Is Octopi open source?
No. Octopi is a free app, not open-source software. The models it runs are open-weight: the default, Qwen2.5-7B-Instruct, is published under the Apache 2.0 license, and other models’ licenses vary.
Related
Run an open-weight model on your Mac
Octopi runs open-weight models locally through Ollama — free, with no account needed.
Download Octopi free