Not all models are used the same way. Two simple questions sort most of them: can you download it? and where does it run?
Open and closed models
- Closed (proprietary) models are only available through the company's app or API. You can use them, but you can't download them or see their parameters. Examples: Gemini, GPT, and Claude models.
- Open-weight models publish their trained parameters (weights), so anyone can download them and run them on their own computer or server, within the licence terms. Examples: Gemma, Llama, and Mistral's open models.
| Closed models | Open-weight models | |
|---|---|---|
| How you use them | Through the provider's API | Download and run yourself, or use a hosting service |
| Strengths | Usually the most capable; nothing to set up | Run privately, offline, customise, no per-request fee to a provider |
| Trade-offs | Data leaves your system; pay per use | You need suitable hardware and must run and update them yourself |
"Open-weight" doesn't always mean everything is open — the training data and code may not be published, and licences differ. Always read the licence before using one in a product.
Cloud and on-device
- Cloud models run on powerful servers. Your app sends a request over the internet and waits for the answer. This is how the largest, most capable models are used.
- On-device models run directly on the phone or laptop. They are much smaller, so they are less capable — but they work offline, respond without a network round trip, and keep data on the device.
On Android, Gemini Nano is an example of an on-device model available on supported phones, and smaller open models like Gemma can also be run on device with the right tools.
Cloud: phone ──internet──▶ big model on a server ──▶ answer
On-device: phone ──▶ small model on the phone ──▶ answer (no internet needed)
How to choose
Ask three questions about your feature:
- How hard is the task? Complex reasoning or long documents → cloud. Short summaries, smart replies, simple rewriting → on-device may be enough.
- How private is the data? Very sensitive data that shouldn't leave the device → on-device or self-hosted open model.
- Does it need to work offline or instantly? → on-device.
Many real apps combine both: try the on-device model first, and fall back to the cloud for harder requests.
Quick check: you're building a feature that suggests short replies to messages, and it must work without internet. Cloud or on-device? (On-device — the task is simple and it must work offline.)