Reference
Find your RAM and graphics memory, then estimate what a model needs.
1 · Find your memory on Windows or macOS
Windows
Open Task Manager → Performance → Memory. Note the installed RAM and how much is available with your usual apps open. Then select your GPU and look at its dedicated GPU memory and its shared GPU memory separately.
Schematic · not your device
- Memory
- Installed / in use / available
- Dedicated GPU memory
- The graphics card’s own capacity
- Shared GPU memory
- Uses system RAM; not extra RAM
Names vary by Windows version and hardware. Microsoft’s GPU-memory explanation
macOS
Choose Apple menu → About This Mac to see the chip and installed memory. Apple silicon Macs use unified memory. Then open Activity Monitor → Memory and, with your usual apps running, look at memory used, memory pressure and swap.
Schematic · not a measurement
- Physical memory
- Installed capacity
- Memory pressure
- How well memory is keeping up with demand
- Swap used
- Disk space the system is using as overflow for memory
Free memory alone doesn’t tell you whether your Mac is coping; memory pressure does. An Intel Mac can have a different graphics-memory layout. Apple’s memory guide
RAM and VRAM are working memory. Free disk space isn’t extra RAM, so leave it out of the total.
Instructions checked 16 September 2026.
2 · Check a model file before you download it
- Read the publisher’s model card for the task, the variant and the license. It may not say which apps can run the model, so check your runner’s support separately.
- In your runner’s catalog, find a supported file and see who published or converted it. Make sure the model, precision and format are the ones you want.
- Look for companion files the model needs, and compare the download size with your free disk space. The download size isn’t the memory the model will need to run.
- Use the guide’s weights calculation to shortlist candidates.
For example, Google’s Gemma 3 12B instruction-tuned model card tells you about the source model. A converted file for your runner is a separate download, and its support has to be checked on its own.
3 · Estimate a downloaded model with LM Studio
This example uses LM Studio’s command-line tool, lms. Install LM Studio from its official site and follow its CLI setup instructions. Then, in its catalog, pick a model file your installed runtime supports and download it. The commands below only inspect that local file; they don’t download anything.
- Find your model’s key.
lms lsCopy the identifier shown for your downloaded file. If the command isn’t found, finish the official CLI setup first.
- Ask for an estimate with settings you choose.
lms load --estimate-only "MODEL_KEY" --context-length 4096 --gpu offReplace
MODEL_KEYwith the key you copied. This example is CPU-only with a 4,096-token context; it’s a demonstration, not a setting to use for every task. To see a GPU configuration, use a supported GPU setting and run the estimate again. - Read the estimate before loading.
Output fields explained · no real result shown
- Estimated GPU Memory
- Compare it with the memory the GPU can reach in this configuration.
- Estimated Total Memory
- The runner’s estimate for the model’s whole runtime. It doesn’t include your operating system or your other apps.
- Loadability message
- A resource check, not a promise of useful speed or correct answers.
Field names can differ between runtimes and versions. Don’t add the total and GPU figures together; on a unified-memory machine both draw on the same installed capacity. Run the estimate again whenever you change the context, the offload setting or the model file.
If the estimate leaves enough room, load the model in LM Studio with the same settings and try your actual task, watching memory pressure and how responsive it feels. If it doesn’t fit, shrink the workload or choose a smaller model your runner supports.
Checked on 16 September 2026 against LM Studio’s load and estimate documentation. Other runners may not have an equivalent. If yours doesn’t, go by its documented requirements, start with a small supported model and watch memory use.
4 · Find your memory on Linux with an NVIDIA GPU
Open a terminal. nvidia-smi shows the GPU’s name and its memory, and free -h shows system RAM. The GPU’s memory is separate from system RAM.
- GPU name and memory.
nvidia-smiLook for the product name and the memory used and total for each GPU. That’s the card’s dedicated video memory, its VRAM.
- System RAM.
free -hIn the
Memrow, read the total and available columns. Swap is disk space, so it isn’t extra RAM.
Schematic · not your device
- GPU memory (VRAM)
- The card’s own capacity, reported by
nvidia-smi - System RAM
- Installed memory, reported by
free -h - Available
- An estimate of memory that can start new work without swapping
Spare system RAM doesn’t add to the card’s VRAM. Compare the weights with whichever pool your runner will actually use.
Linux commands checked on 23 October 2026 against NVIDIA’s System Management Interface documentation and the free manual page. Output layout differs by driver and distribution.