Open-Source vs Open-Weight: what's the difference and which Ollama models are which?
The rise of publicly accessible AI models has given birth to two major trends: open-source and open-weight. While often used interchangeably, they represent very different realities. Open-source promotes full transparency — code, weights, training data, and documentation. Open-weight offers partial openness, limited to the model's learned parameters.
As the Open Source Initiative (OSI) has made clear, models like Meta's Llama are not open-source under the official definition. They are open-weight — downloadable and usable, but lacking the full transparency required by the OSI's Open Source Definition.
This distinction matters more than ever for developers, enterprises, and governments seeking AI sovereignty. When you run a model via Ollama, understanding its license determines what you can legally do with it in commercial products.
Open-Source: total openness
A truly open-source AI model provides:
- The full source code used to build and train the model
- The model weights (parameters)
- The training data (or detailed information about it)
- Complete documentation for understanding and modifying the model
OSI-approved licenses like Apache 2.0 and MIT grant unrestricted commercial use, modification, and redistribution.
Open-Weight: partial openness
An open-weight model makes only its weights available — the parameters learned during training. The training data, design code, and methodology remain proprietary. This means:
- The model can be used and fine-tuned, but not fully inspected
- Biases or weaknesses are harder to analyze
- Adoption is easier, but control stays with the original developer
Most models marketed as "open-source" in AI are actually open-weight.
Categorizing Ollama models
Below is a list of popular models available on Ollama, classified by their true license type.
Truly open-source models (Apache 2.0 or MIT)
These models use OSI-approved licenses and grant full commercial freedom:
| Model | Developer | License |
|---|---|---|
| Qwen 3, Qwen 3.5, Qwen 2.5 | Alibaba | Apache 2.0 |
| DeepSeek V3, R1, R2, V4 | DeepSeek | MIT |
| Mistral 7B, Mistral Small | Mistral AI | Apache 2.0 |
| Gemma 4 | Apache 2.0 | |
| Phi-3, Phi-4 | Microsoft | MIT |
| GPT-OSS (20B, 120B) | OpenAI | Apache 2.0 |
| OLMo | AI2 (Allen Institute) | Apache 2.0 |
| IBM Granite (3B–32B) | IBM | Apache 2.0 |
| Nomic Embed Text | Nomic | Apache 2.0 |
| FLUX.1 Schnell | Black Forest Labs | Apache 2.0 |
Open-weight models (restrictive / custom licenses)
These models provide downloadable weights but carry usage restrictions:
| Model | Developer | License | Restriction |
|---|---|---|---|
| Llama 3.1, 3.2, 3.3, 4 | Meta | Llama Community License | 700M MAU threshold |
| Mistral Large, Codestral | Mistral AI | Custom Mistral License | Research / non-commercial |
| Gemma 2, Gemma 3 | Gemma License | Use restrictions | |
| Command R, Command R+ | Cohere | CC-BY-NC | Non-commercial only |
| Kimi K2.6, K2.7 | Moonshot AI | Modified MIT | Additional restrictions |
| MiniMax M3, M2.5 | MiniMax | Custom | Usage restrictions |
| GLM-4, GLM-5 | Zhipu AI | Custom | Usage restrictions |
| DeepSeek V2 | DeepSeek | Custom | Older restrictive license |
Why this matters
For developers
Choosing a model with a permissive license (Apache 2.0 or MIT) means you can build, fine-tune, deploy, and sell your product without legal uncertainty. Open-weight models with custom licenses may restrict commercial use, impose revenue-sharing, or limit deployment scale.
For enterprises
Legal teams review licenses carefully. Apache 2.0 provides explicit patent protection. MIT is the simplest to comply with. Custom licenses require case-by-case legal review and may block deployment entirely.
For AI sovereignty
Governments and European institutions seeking independence from US tech giants should prioritize truly open-source models. Only Apache 2.0 and MIT-licensed models provide the durable freedoms needed for sovereign AI infrastructure.
How to check a model's license in Ollama
You can verify any model's license directly from the command line:
ollama show <model> --license
For example, to check Llama 4's license:
ollama show llama4 --license
The bottom line
The AI industry is flooded with "open-washing" — marketing open-weight models as open-source. Before building your next project, check the license. If it's Apache 2.0 or MIT, you have real open-source freedom. If it's a custom license, you're using an open-weight model with strings attached.
The good news: truly open-source models like Qwen 3.5, DeepSeek R2, Gemma 4, and GPT-OSS are now competitive with proprietary alternatives, and you can run them all locally with Ollama.