Answer first: Choose RAM for local LLMs by adding model weights, context or KV-cache memory, runtime overhead and operating-system headroom. A 32 GB system is a practical experimentation tier, but it is not a universal recommendation; the model and context decide the requirement.
Last reviewed: October 13, 2026. Mini PC Land has not hands-on tested every configuration discussed here. Recommendations are based on documented platform capabilities, official compatibility information and workload requirements. Confirm the exact model and configuration before buying.
← Explore all mini PC use cases · How we research
- Match capacity to the real workload.
- Verify the exact processor, ports and upgrade limits.
- Check software support before buying for acceleration.
- Keep backups and a recovery path.
Why parameter count is not enough
A model with billions of parameters can be stored at different precisions. Quantization reduces weight memory, but runtime buffers and the key-value cache remain. Longer context, larger batches and concurrent sessions increase memory use. Published model file size is therefore a starting point, not the final requirement.
A safe estimation method
Begin with the actual model file size. Add the runtime estimate for the chosen context length and number of sessions. Add memory used by the operating system and application. Then keep at least 15 to 25 percent free so the machine does not rely on slow disk swapping. Validate with the runtime before purchasing non-upgradeable hardware.
System RAM, VRAM and unified memory
On a CPU-only system, model data lives in system RAM. A discrete GPU is constrained by VRAM unless the runtime can split layers between GPU and CPU, which reduces the hard failure risk but may reduce speed. Integrated graphics borrow system memory. Unified-memory systems share one pool but still reserve capacity for the OS and graphics.
Practical capacity tiers
Sixteen gigabytes can run small quantized models but leaves little flexibility. Thirty-two gigabytes is a useful general experimentation tier. Sixty-four gigabytes opens larger quantized models and longer contexts, assuming memory bandwidth and compute remain acceptable. Ninety-six or 128 GB targets users who already know why they need it; capacity alone does not make inference fast.
Two upgrade traps
First, confirm the maximum supported memory for the exact mini PC, not only the CPU specification. Firmware and slot layout matter. Second, some integrated-GPU applications need configurable shared memory or specific drivers. Do not assume all installed RAM can be allocated to the accelerator.
Test before you commit
Run your intended runtime and model on any available machine, or consult measurements with the same quantization and context. Record peak memory during model load and generation. That measurement is more valuable than a generic “models per gigabyte” chart.
Use this with the local AI buying guide and NPU, iGPU and eGPU comparison.
Frequently asked questions
Is 16 GB enough for a 7B model?
Some quantized models can fit, but context, runtime overhead and the operating system reduce available memory. Thirty-two gigabytes gives more margin.
Does longer context use more RAM?
Yes. The key-value cache generally grows with context length and can materially change the requirement.
Can RAM replace GPU VRAM?
Some runtimes can offload part of a model to system RAM, but the slower connection and memory bandwidth usually reduce performance.
Sources and update notes
This guide uses official documentation where a platform capability is stated. Compatibility changes, so follow the linked source and verify the current release before purchase or deployment.