How to Use a Local LLM on an 8GB RAM Laptop: Best Models and Settings

Yes, an 8GB RAM laptop can run a local LLM, but the useful range is small models, short context windows, and one model at a time. The goal is not to squeeze the biggest model into memory. It is to keep enough RAM free for Windows, macOS, or Linux so the machine stays responsive while the model is running.

What Actually Fits on 8GB RAM?

On an 8GB machine, the operating system, browser, antivirus, and background apps already take a meaningful share of memory. That is why 1B–3B-class models are the practical starting point. A 4B model can work in some setups, but it leaves less headroom and becomes much more sensitive to context length and background apps.

Model download size is not the same as total RAM use. Ollama also needs runtime memory and a KV cache for the conversation. Ollama notes that increasing context length requires more memory, so a model that technically loads can still become sluggish in a long chat.

Model	Current Ollama size	8GB laptop use
Gemma 3 1B	815MB	Very light; useful for short summaries, rewriting, and simple Q&A
Llama 3.2 1B	1.3GB	Safe starting point for general tasks
Qwen3 1.7B Q4_K_M	1.4GB	Good balance when you want more capability without a large footprint
Llama 3.2 3B	2.0GB	Stronger general chat; realistic if you keep other apps closed
Phi-4 Mini 3.8B	2.5GB	Stretch option; use a short context and plenty of free RAM

Those sizes come from Ollama’s current model library. They are useful for comparing model footprints, not as a promise of total system memory use.

How to Run a Local LLM on an 8GB Laptop

1. Free RAM before loading the model

Close memory-heavy browser tabs, game launchers, creative apps, and anything else you do not need. If you are buying a machine specifically for local AI, this laptop buying guide explains why RAM matters; the budget laptop list under ₹40,000 is also useful if 8GB is your current ceiling.

2. Install Ollama

Download it from ollama.com. Ollama now has a desktop app on Windows and macOS, so you are not limited to a terminal-only workflow. If you want the complete setup, use our Ollama laptop guide.

3. Start with a small model

For the safest first run, try:

ollama run qwen3:1.7b

If that runs comfortably, try ollama run llama3.2:3b. If the machine starts swapping heavily or becomes difficult to use, step back to a 1B–1.7B model rather than forcing a larger one.

4. Keep context short

A model may advertise a 32K, 128K, or larger context window, but that does not mean an 8GB laptop should use it. Start around 2K–4K tokens. Longer context increases memory use, and that extra memory sits on top of the model itself. For low-RAM hardware, a fresh chat is often better than carrying one conversation for dozens of turns.

5. Use one local-AI stack at a time

Do not keep models loaded in Ollama and LM Studio at the same time on an 8GB machine. If you prefer a browser-style interface, install Open WebUI on Windows and connect it to Ollama instead of loading another model runtime.

Settings That Matter More Than People Expect

• Quantization: Prefer Q4 variants on this class of hardware. Higher-precision versions consume more memory.
• Context: Begin at 2048–4096 tokens. Increase only if the machine remains stable.
• Swap/page file: Keep it enabled. It is slower than RAM, but it can prevent a hard crash during a memory spike.
• Power mode: Plug the laptop in when possible. Aggressive battery-saving modes can cut CPU performance.
• Loaded models: One at a time. On a low-memory system, headroom matters more than convenience.

If It Is Slow or Freezing

If the model loads but the whole laptop becomes unresponsive, the first fix is usually smaller—not more tweaking. Drop from 3B/4B to 1B–1.7B, shorten the context, close the browser, and restart the session. Ollama’s newer scheduling engine measures model memory requirements more accurately than older releases, but it cannot create RAM that the machine does not have.

Avoid trusting universal “tokens per second” numbers for low-end laptops. CPU generation speed changes dramatically with processor generation, memory bandwidth, power limits, and thermal throttling. A better test is whether your own machine can answer at a pace you are comfortable reading without making the rest of the system unusable.

Best 8GB Setup to Start With

For most 8GB laptops, start with Ollama plus Qwen3 1.7B or Llama 3.2 1B, keep context at 2K–4K, and close heavy background apps. Move to Llama 3.2 3B only if the smaller model runs comfortably. That setup is enough for private rewriting, summaries, simple coding questions, notes, and lightweight offline assistance.

If the laptop is not available, you can also run AI locally on Android without internet. The same principle applies: a smaller model that stays responsive is more useful than a larger model that only barely fits.