Local AI VRAM calculator: what can my PC run?
Enter how much memory your graphics card has and this calculator shows which local AI model sizes our VRAM guide recommends for it at 4-bit quantization, the compression most local tools use by default. For most people 8 GB of VRAM is the comfortable starting point, and 12 to 16 GB covers the popular mid-size models. All figures are approximate.
Check your hardware
On Windows: Task Manager, Performance, GPU, Dedicated GPU memory.
With 8 GB of VRAM, the guide's table puts models up to about 7 to 8 billion parameters within range at 4-bit quantization.
This estimates memory for one chat model. It does not check GPU compatibility or InnerZero's full setup, which the setup wizard handles when you install. InnerZero's defaults are deliberately conservative, so setup may pick a smaller model than the largest size shown here.
- 1 to 2 billion parametersWithin the guide's range
Guide recommends 2 to 4 GB or none
Catalogue examples: Gemma 3 1B
- 3 to 4 billion parametersWithin the guide's range
Guide recommends 4 to 6 GB
Catalogue examples: Qwen 3 4B, Gemma 3 4B
- 7 to 8 billion parametersWithin the guide's range
Guide recommends 6 to 8 GB
Catalogue examples: Qwen 3 8B
- 12 to 14 billion parametersBelow the guide's recommendation
Guide recommends 10 to 12 GB
Catalogue examples: Qwen 3 14B, Gemma 3 12B
- 27 to 32 billion parametersBelow the guide's recommendation
Guide recommends 24 GB
Catalogue examples: Qwen 3 30B, Qwen 3 32B
- 70 billion parametersBelow the guide's recommendation
Guide recommends 48 GB or multi-GPU
No catalogue examples in this size
A model a little larger than your VRAM may run partly in system RAM, more slowly, if you have enough RAM and your backend supports it.
VRAM needed by model size
The full table from our guide to how much VRAM you need for local AI, with examples of each size from our model catalogue.
| Model size | Approx 4-bit size | Recommended VRAM | Runs on CPU only? | Catalogue examples |
|---|---|---|---|---|
| 1 to 2 billion | About 1 GB | 2 to 4 GB or none | Yes, easily | Gemma 3 1B |
| 3 to 4 billion | About 2 to 3 GB | 4 to 6 GB | Yes, usable | Qwen 3 4B, Gemma 3 4B |
| 7 to 8 billion | About 4 to 5 GB | 6 to 8 GB | Yes, slower | Qwen 3 8B |
| 12 to 14 billion | About 8 to 9 GB | 10 to 12 GB | Possible, slow | Qwen 3 14B, Gemma 3 12B |
| 27 to 32 billion | About 18 to 20 GB | 24 GB | Hard, very slow | Qwen 3 30B, Qwen 3 32B |
| 70 billion | About 40 GB | 48 GB or multi-GPU | Not practical | None in the catalogue |
Running on CPU only
With no graphics card the model runs on your CPU using system RAM, more slowly than on a GPU, so system RAM sets the limit. These are the working recommendations from our guide to running local AI without a GPU.
| Your RAM | Realistic model size | What you can do |
|---|---|---|
| 8GB | 1B to 4B (4-bit) | Chat, writing help, summaries, simple questions |
| 16GB | 7B to 8B (4-bit), or gpt-oss 20B | Solid daily assistant, light coding help |
| 32GB+ | 8B comfortably, larger when patient | Headroom for big context and multitasking |
Below 8GB you can still run a tiny model, but keep your prompts short to avoid the system swapping to disk. InnerZero needs at least 4 GB of RAM to run a local model.
Quantization starting points
Separate examples from our guide to LLM quantization. They apply to graphics card memory, and each one names its own model size and quantization level.
- 8 GB of VRAM
- A 7B to 8B model at Q4_K_M
- 12 GB of VRAM
- An 8B model at Q5_K_M or Q6_K, or a 14B at Q4_K_M
- 16 GB of VRAM
- A 14B model at Q4_K_M to Q5_K_M comfortably
- 24 GB of VRAM or more
- A 14B at Q8_0, or a larger model at Q4_K_M
How this calculator works
It compares your VRAM with the lower end of the recommended range for each model size in our VRAM guide. As a rule of thumb, a 4-bit model needs about 0.5 GB of memory per billion parameters for its weights, plus 1 to 2 GB of headroom for the context window. Context length and the exact quantization level shift the real figure up or down, and moving up to 8-bit roughly doubles the memory a model needs.
The figures cover one chat model. A complete assistant also needs a small embedding model for memory and, on capable hardware, local voice models. InnerZero's setup wizard checks your hardware and picks the models to match, so you do not have to add this up yourself.
Frequently asked questions
How many parameters can my GPU handle?
At 4-bit quantization, our VRAM guide recommends 2 to 4 GB or none for 1 to 2 billion parameters, 4 to 6 GB for 3 to 4 billion, 6 to 8 GB for 7 to 8 billion, 10 to 12 GB for 12 to 14 billion, 24 GB for 27 to 32 billion, and 48 GB or multi-GPU for 70 billion. All figures are approximate.
Is 8 GB of VRAM enough for local AI?
For most people, yes. 8 GB is the comfortable starting point, and the guide's table puts popular 7 to 8 billion parameter models within range at 4-bit quantization. The 12 to 14 billion parameter size wants 10 to 12 GB.
Is 16 GB of VRAM enough for local AI?
For models up to 12 to 14 billion parameters, yes. Our quantization guide suggests a 14B model at Q4_K_M to Q5_K_M runs comfortably on 16 GB. The 27 to 32 billion parameter size wants 24 GB.
Does local AI use RAM or VRAM?
Both, depending on your hardware. With a graphics card the model loads into VRAM, which is fastest. With no graphics card, system RAM holds the model and the CPU does the work, which is slower. A model a little larger than your VRAM may run partly in system RAM, more slowly, if you have enough RAM and your backend supports it.
What can I run with no graphics card?
With no graphics card the model runs on your CPU using system RAM, more slowly than on a GPU. Our guide to running local AI without a GPU gives these realistic model sizes. 8GB of RAM: 1B to 4B (4-bit). 16GB: 7B to 8B (4-bit), or gpt-oss 20B. 32GB+: 8B comfortably, larger when patient. These are general examples, not a list of the models InnerZero installs.
How accurate is this VRAM calculator?
It gives approximate planning figures, not exact numbers. Context length, the exact quantization level and the model family all shift real memory use. It estimates one chat model and does not check GPU compatibility or InnerZero's full setup. A model uses roughly its download size plus 1 to 2 GB once loaded, so check the listed size before a big download.
Let InnerZero pick for you
InnerZero's setup wizard checks your hardware, tells you what it found and picks the models to match, so you do not have to work out the VRAM maths yourself. On a graphics card, its defaults leave room for voice, speech recognition and the context window alongside the chat model, and you can pick a different hardware profile yourself from Manual Profiles in Settings. InnerZero is free to download on Windows, macOS, and Linux, with no account required.