GET STARTED WITH LOCAL AI
How to check your VRAM and RAM for local AI
You need two numbers for a PC: the memory on your graphics card and your system RAM. An Apple Silicon Mac uses one shared memory pool. Once you have those figures, our finder can estimate which models fit.
How to check VRAM and RAM on Windows
- Press Ctrl + Shift + Esc to open Task Manager, then select Performance.
- Select your dedicated GPU. If there is more than one, check the graphics card you intend to run the model on.
- Find Dedicated GPU memory. A reading such as 1.0 / 8.0 GB means roughly 1 GB is in use out of 8 GB total. Do not add shared GPU memory to this figure.
- Select Memory to see your system RAM. This is separate from your GPU memory.
Choose Windows / Linux PC (dedicated GPU) in the finder. Start with the capacity figures, then allow for other applications using memory. The calculator already includes a loading allowance and GPU headroom, but those allowances cannot cover every workload.
Microsoft explains the distinction between dedicated and shared memory in its Task Manager GPU guide.
How to check memory on an Apple Silicon Mac
Open the Apple menu and choose About This Mac, or open System Settings → General → About. Look for the chip and memory fields. An Apple M-series chip identifies an Apple Silicon Mac.
Choose Apple Silicon Mac in the finder and enter its unified memory once. Do not count it separately as both GPU memory and system RAM. The Mac target includes an allowance for macOS and other memory use. Intel Macs are not covered by this mode.
How to check memory on Linux
Your desktop's system information or system monitor usually shows RAM. With a supported NVIDIA driver installed, running nvidia-smi in a terminal shows GPU memory use and capacity. Divide MiB by 1,024 to get GiB: 8,192 MiB is 8 GiB.
For AMD or Intel graphics, use the vendor's system information tools to check the exact GPU and its dedicated memory. Integrated graphics and CPU-only inference are outside this finder's current scope. Do not combine the capacities of multiple GPUs.
Which local AI settings should I choose?
Runtime: the app that runs the model
Select the software you intend to use. LM Studio provides a desktop interface for local models; Ollama supports local model workflows through its app and tools. vLLM is aimed at serving models and requires more setup. Check the runtime's hardware support before downloading: enough memory alone does not prove compatibility.
Official starting points: LM Studio documentation and Ollama documentation.
Context: how much the model can keep in one conversation
Context includes the instructions, chat history, supplied documents and generated response. Tokens are pieces of text, not necessarily whole words. For a first memory comparison, try 4,096 tokens, then increase it if your task needs longer documents or conversations. This is a starting point, not a guarantee that every task will fit.
A longer context generally needs more cache memory. The finder excludes models whose supported context is shorter than your selection. Set the same context in your runtime when you test it. Ollama explains how context affects memory.
Weight format: smaller downloads or more precision
Balanced (Q4 / MXFP4) is a practical first comparison for Ollama or LM Studio. Quantization stores weights more compactly, usually reducing memory needs, with a possible quality trade-off. Q8 uses more memory than Q4; 16-bit weights usually need more again. More precision does not guarantee a better answer for your task.
gpt-oss uses its native MXFP4 format. For vLLM, this calculator supports 16-bit estimates and native gpt-oss MXFP4 where the backend keeps the expert weights compressed. It does not treat all formats as interchangeable.
What does a matching model mean?
A match fits our estimated memory target at your chosen context and format. For a PC, both VRAM and system RAM must fit. For a Mac, the unified-memory target must fit. Results are alphabetical, not a quality ranking.
Memory left is your entered capacity minus the target, which already includes headroom. Use Explore to transfer the exact model, format, runtime and context to the detailed calculator. Check its architecture notes and actual model file before installing.
What if no models fit?
Try a shorter context or a smaller weight format. Check you have not entered storage capacity in place of memory. No match means none of our catalog entries fit this calculation; it does not mean your computer cannot run any AI. CPU offload, multiple GPUs, image-processing workloads and CPU-only setups are not modelled.
Will a model that fits be fast?
Not necessarily. Memory capacity is only one constraint. Speed also depends on memory bandwidth, compute, runtime support and the workload. These are planning estimates, not measured benchmarks.