Quick overview
Ollama PC requirements for Windows and Linux: GPU VRAM, RAM, disk space, and how to validate a model before you upgrade hardware.
Ollama lowers the friction of running open models on a desktop, but it does not repeal VRAM limits. Requirements come from the model tag you pull, the quant baked into that tag, and how much context you configure. This page translates Ollama usage into hardware decisions for Indian home labs and small offices.
Quick answer
Check each model's listed size in `ollama pull` output. Add RAM and disk headroom for downloads and concurrent apps. GPU acceleration needs a card Ollama can drive with enough VRAM for weights plus context. CPU-only mode is possible but slow for larger models.
What Ollama checks on your machine
- Disk: model blobs under the Ollama storage path
- GPU VRAM: when layers stay on GPU
- System RAM: CPU inference or partial offload
- CPU cores: affect tokens per second when GPU fit is partial
Practical validation workflow
1. Pull the smallest quant that meets your quality bar. 2. Run a 10-minute session at your real context length. 3. Watch for OOM errors or thermal throttling. 4. Upgrade hardware only after you reproduce the bottleneck.
Compare with LM Studio
Different UI, similar physics. Read LM Studio vs Ollama hardware needs.
Link to broader AI PC concepts
If you are new to local inference, start with what is an AI PC.
How Roger N Reckon can help
Need a machine pre-tested with your Ollama model list? Roger N Reckon can document load tests during workstation delivery.
Practical checklist
1. Write down the model size, quant level, and target context length you need. 2. Measure current GPU VRAM, system RAM, and free NVMe space before buying parts. 3. Confirm PSU wattage against the GPU vendor's required system power. 4. Plan cooling for sustained inference, not just short benchmarks. 5. Back up data before OS or hardware changes. 6. Request a written component list, testing scope, and warranty terms for any build or upgrade. 7. Re-test with the same prompt and context after changes.
Frequently Asked Questions
Can Ollama run without a GPU?
Yes, on CPU, but large models become impractical. Small 3B–7B models are the realistic CPU-only range for interactive use.
How much disk for Ollama models?
Plan tens of gigabytes per large model. Keep 20% free space on the drive for updates and temporary files.
Does WSL change requirements?
WSL2 GPU passthrough adds complexity. Native Windows or Linux is simpler for troubleshooting thermals.
Why does the same 8B tag fail on two PCs?
Different quants, context sizes, or background GPU usage (browsers, monitors) change free VRAM.
Should businesses run Ollama on office PCs?
Only with IT policy, patch cadence, and backups. Sensitive prompts still need access control.
Final takeaway
Ollama is an installer and runtime, not a VRAM upgrade. Size hardware to the tags you actually pull.