Can You Run Local AI on a PC with 8GB of RAM? How to Choose the 'Just Right' Model Based on Real Measurement Data
Have you given up, thinking, 'It's impossible to run AI with my PC specs'? Actually, even with 8GB of RAM, it is possible to run AI surprisingly smoothly if you choose the appropriate model size.
- Even after subtracting the portion used by the OS (approx. 4GB), models under 4GB are well within the practical range
- Lightweight models (1B–2B class) produce answers in an instant even with CPU only
- 3B–4B class models achieve a 'comfortable' feel, exceeding the speed at which humans can read
- With a GPU (video card), you can expect speeds several to dozens of times faster
The 'Prerequisites' for Running on a PC with 8GB RAM
When running local AI (AI run directly on your own PC), the most important factor is the amount of free 'memory' (RAM). A PC's memory consumes a certain amount from the start to run the OS (Windows, Mac, etc.) and browsers.
In this case, I investigated models that can run with the remaining memory, assuming that approximately 4GB of memory is already being used for the OS and other processes. For a PC with 8GB of RAM, the calculation suggests that if you choose a model where the memory used is 4GB or less, it can be operated within a practical range.
- Basic OS operation: Consumes approx. 4GB (assumed)
- Operational benchmark: If the model's memory usage is 4GB or less, it is possible to operate even on a PC with 8GB of RAM
[Real Measurements] Operating Speed and Memory Usage by Model
When I actually ran multiple models, I found that there are clear differences in memory consumption and generation speed depending on the model size (number of parameters). The following figures are for cases where it was run on CPU only.
Here, 'tokens' refer to the minimum unit of text an AI handles. For Japanese, this corresponds to roughly 1–2 characters. We judge the perceived level of comfort based on the speed of these tokens (the amount produced per second).
- gemma3:1b (extremely lightweight)
- - Memory used: 0.9GB
- - CPU speed: 61.7 tokens/sec (level where characters appear instantly)
- llama3.2:3b (balanced type)
- - Memory used: 2.3GB
- - CPU speed: 27.3 tokens/sec (faster than reading, very comfortable)
- qwen3.5:2b (lightweight / high efficiency)
- - Memory used: 2.3GB
- - CPU speed: 27.7 tokens/sec (faster than reading, very comfortable)
- gemma3:4b (slightly larger)
- - Memory used: 2.8GB
- - CPU speed: 21.5 tokens/sec (comfortable speed, roughly faster than reading)
- qwen3.5:4b (slightly larger)
- - Memory used: 3.1GB
- - CPU speed: 19.3 tokens/sec (speed that operates comfortably)
Conclusion: This is the Correct Way to Choose for 8GB RAM
The point revealed by the real measurement data is that even in an 8GB RAM environment, you can build a stress-free AI environment depending on your 'model selection.'
In particular, for models with memory usage around 3GB, you can obtain answers at a speed that exceeds the human reading speed even when running on CPU only. This is more than comfortable enough to use as an assistant for daily tasks.
If you seek even faster operation (such as generating text instantly), the presence of a GPU (a part dedicated to image processing) becomes important. If you have a GPU, these figures will jump by several to dozens of times, making it easier to handle more advanced tasks.
- If seeking the fastest: gemma3:1b (instant even on CPU)
- If prioritizing balance: llama3.2:3b or qwen3.5:2b (comfortable feel)
- If trying slightly smarter models: gemma3:4b or qwen3.5:4b (practical speed)
Summary
Before giving up and thinking, 'It won't run on my PC,' I recommend trying lightweight models first. Even with the constraint of 8GB of RAM, if you choose the appropriate model, the experience of running AI in a local environment is fully possible.
First, check how many GB of memory your PC has, and start with the 'just right size' models introduced here.
- Understand your PC's specs
- Check the model size (memory usage)
- Start by trying models around 3GB
- People who have a PC with 8GB of RAM and are anxious about whether they can run local AI
- People who want to know how much they can do with their PC specs
- People who want to try AI in their own environment without relying on paid services
- Local AI
- Artificial intelligence that operates directly on your own PC hardware without using cloud services (like ChatGPT).
- Model
- A collection of AI training data and structures. It is essentially the 'AI's brain,' and its intelligence and operating speed vary depending on its size (number of parameters).
- Token
- The minimum unit when AI processes text. In Japanese, this corresponds to approximately 1–2 characters.
- CPU
- The primary brain component that performs calculations on a computer. It excels at general-purpose processing.
- GPU
- A component that excels at image processing and massive parallel calculations. It plays an important role in accelerating AI inference.
FAQ
Can it really run comfortably with 8GB of RAM?
Yes, it runs comfortably if you choose the appropriate model size. In this measurement, it was confirmed that models around 3GB can operate at a 'faster than reading' comfortable speed even on CPU only.
What happens if I don't have a GPU?
Even without a GPU, lightweight models (1B–4B class) can be run at practical speeds using only the CPU. However, the speed will be more gradual compared to when a GPU is present.
Even on a PC with 8GB of RAM, you can run local AI at a sufficiently practical speed by choosing the appropriate model size (specifically those under 4GB). Real measurement data showed that models around 3GB can maintain a 'faster than reading' comfortable speed even on CPU only. The key to a smooth introduction is to start by trying lightweight models.