PaperLabCan you run an AI model locally for research work?
Yes. Tools like Ollama and LM Studio run open language models on your own laptop, so papers, drafts and data never leave it and each question costs nothing. A laptop with 16 GB of memory runs models of about 7–8 billion parameters comfortably. They answer questions about a supplied passage well, but are weaker than the largest cloud models at long reasoning, so give them the right passages and check their answers.
By Malek Khannoussi · Updated25 September 2026
What you need
| Memory | What runs well |
|---|---|
| 8 GB | Small models (3–4 billion parameters): short answers, simple questions |
| 16 GB | 7–8 billion parameter models: good answers from supplied passages |
| 32 GB or more | Larger models, longer context, faster answers |
Set it up in three steps
- Install Ollama.
- Pull a model, for example
ollama pull qwen3:8b. - Connect an app that sends your papers' passages to it, and cites them.
Make a small model do well
- Give it the right passages: retrieval matters more than model size.
- Ask one clear question at a time.
- Ask for the passage it relied on, and check it.
- Use a cloud model only for what the local one can't do, and only on text you can share.
New to the terms? See local LLM andembeddings in the glossary.
Questions people ask
Which local model should I start with?
A recent 7–8 billion parameter instruction model, such as a Qwen or Llama variant served by Ollama. Try a few on questions you know the answers to.
Do I need a GPU?
No. Apple silicon Macs run local models well on their unified memory, and modern CPUs work, more slowly. A GPU makes answers faster.
Is a local model good enough for research?
For answering from passages you provide, often yes. For long reasoning without sources, the largest cloud models are still stronger.
Is it really private?
The model runs on your machine, so your text isn’t sent anywhere by it. Check that the app around it doesn’t send data elsewhere either.