The short answer
There is no single RAM or VRAM requirement for every Ollama model. Budget for weights, conversation cache and runtime buffers, then leave room for the system. An 8GB GPU can be useful for smaller quantized models; 12–16GB provides more room, and 24GB can support larger workloads. Check the exact model and context before treating any tier as sufficient.
RAM, VRAM and disk space do different jobs
On a PC with a discrete GPU, VRAM is the memory on the graphics card. System RAM belongs to the rest of the computer. A machine with 32GB RAM and an 8GB GPU does not have a 40GB GPU. CPU offloading can make a larger model runnable, but it changes where computation happens.
Disk space stores your downloads. It is a separate budget: two quantizations of a model may require separate files, and neither download size tells you the complete live allocation. On Apple Silicon, CPU and GPU use shared unified memory, so macOS and your other apps compete with inference for the same pool.
Start by recording your exact GPU, its memory capacity, system RAM and available storage. Then choose a model artifact and an initial context. This turns a vague hardware question into a configuration you can check.
What can 8GB, 12GB, 16GB or 24GB VRAM run?
Use these tiers to start a shortlist. They are editorial planning ranges for one text conversation, not minimum requirements published by Ollama. Model architecture, the actual quantized file, context and the runtime can move a workload across a tier boundary.
| GPU VRAM | A useful starting experiment | What to watch |
|---|---|---|
| 8GB | A smaller 3–4B model at Q4 and modest context | An 8B model may already be tight once cache and buffers are included. |
| 12GB | An 8B Q4 model; test a 14B Q4 model at short context | A near-fit at 8K may exceed the budget at 32K. |
| 16GB | A 14B Q4 workload with more context headroom | Higher weight precision or parallel requests can use up that headroom. |
| 24GB | Keep a mid-size model resident with more room for context or precision | Larger models still require individual calculation; 24GB is not a universal cutoff. |
Worked examples: weights are only part of the total
These three examples use 99Tokens’ reviewed Ministral 3 architecture records. All use the planner’s Q4 assumption of 4.8 effective bits per stored parameter, FP16 cache, one request and a 1 GiB runtime allowance. They cover text inference; images need additional allocations.
The values are calculated estimates, not measured Ollama usage or exact GGUF sizes. Replace the weight estimate with your actual artifact size in Advanced settings, and verify that your installed runtime supports it.
| Model | Q4 weights | 8K total | 32K total |
|---|---|---|---|
| Ministral 3 3B | 2.2 | 4.0 | 6.4 |
| Ministral 3 8B | 5.0 | 7.0 | 10.2 |
| Ministral 3 14B | 7.8 | 10.0 | 13.8 |
Check the context and GPU placement in Ollama
After loading a local model, run the command below in a second terminal. Read PROCESSOR to see GPU/CPU placement and CONTEXT to see the allocated context. A response appearing in the chat window does not by itself prove the whole model is running on the GPU.
Ollama’s documented defaults currently vary with available VRAM, so explicitly match the context you used in the planner. Begin with a representative short prompt, then repeat with the longest conversation or document you expect to use. Recheck placement after changing settings.
ollama psCompatibility comes before a hardware purchase
Enough memory does not establish GPU support. Check Ollama’s hardware documentation for the exact card, operating system and driver combination. NVIDIA, AMD and Apple use different acceleration paths; a similar product name is not evidence that an unlisted configuration will work.
If you already own the machine, first try a smaller supported model. If it loads but slows down later, inspect context growth and placement before replacing hardware. If the application cannot use the GPU at all, resolve that compatibility problem before evaluating speed.
Common questions before you start
Is 16GB of system RAM enough? It can be useful for smaller local workloads, but it cannot guarantee a particular model will fit. Keep the OS and your ordinary apps in the budget, especially if you plan to offload model work to the CPU.
Do you need a GPU? GPU acceleration is not the same as the ability to run a model. CPU inference is possible, but test whether its latency is acceptable for your task. The calculator assesses memory, not your expected tokens per second.
Will adding another GPU solve it? A compatible runtime must distribute the model across separate devices. Ollama documents single-GPU placement when a model fits, with distribution across available GPUs when it does not. Per-device allocations and the connection between cards still matter.
See it on your own setup.
Change the model, hardware or context and see the memory budget update.
Try this example