Run Llama Locally: Picking the Right Version

By Cedrick Reese, Ready Utilities

Last updated: September 2026

If you followed the beginner guide, you've already run Llama locally: specifically Llama 3.1 8B. This page goes one step further: which Llama version to actually pick, and what to do when something doesn't work.

Which Llama version should you run?

Version	Good for	Notes
Llama 3.1 8B	Getting started, broad compatibility	The most widely supported version across every local AI tool
Llama 3.3 70B	Best quality if you have the memory	Meaningful step up in capability; needs a Mac Studio-class machine or a high-VRAM GPU setup at Q4
Llama 4 Scout	Large context, mixture-of-experts	109B total parameters (MoE, ~17B active per token); the full parameter set still needs to load into memory regardless of how few activate per token

For most people starting out, 3.1 8B is still the right first choice. Check any of these against your exact hardware in the hardware calculator before downloading the bigger versions.

The commands

ollama run llama3.1:8b
ollama run llama3.3:70b

Same pattern as any other model: Ollama downloads it the first time, then starts instantly on every run after that.

Llama-specific troubleshooting

3.3 70B won't load or is extremely slow: this almost always means it doesn't actually fit your available memory. Run the exact number through the calculator rather than assuming "70B should be fine."
Llama 4 Scout uses more memory than you expected for its "active parameter" count: this is expected. Mixture-of-experts models like Scout still need their full parameter set loaded, not just the active fraction. Size your hardware to the total, not the active count.
Responses feel repetitive or oddly formatted: try a different quantization level before assuming the model itself is the problem; very aggressive quantization (well below Q4) can produce this kind of degradation.

Memory figures for Llama 3.3 70B use this site's own calculator formula with real architecture data (80 layers, 8 key-value heads, pulled from the model's published configuration). Llama 4 Scout's mixture-of-experts behavior is confirmed against its official model documentation. This page gets rechecked as new Llama versions ship.

Sources

Frequently asked questions

What's the difference between Llama 3.1 and 3.3?

3.3 is a later, generally stronger version, but at the 70B size, meaning it needs meaningfully more memory than the 8B version most beginners start with. There isn't a 3.3 8B; if you want the smaller size, 3.1 8B is the one to use.

Why does Llama 4 Scout need so much memory if only 17B parameters activate per token?

Because it's a mixture-of-experts model: only a fraction of its parameters compute for any given token, which helps speed, but the entire 109B parameter set still has to be loaded into memory to make that fraction available. Size your hardware to the total, not the active count.

Do I need to uninstall Llama 3.1 before trying 3.3?

No. Ollama keeps every downloaded model separately; you can run either one by name at any time, as long as you have the memory for whichever one you launch.