Published 2026-09-27 · Last updated 2026-09-27
How much RAM do I need for a local LLM?
The short answer
For local LLMs in 2026: 16GB RAM runs 7B–13B models at Q4, 32GB runs up to 32B, and 64GB runs 70B. The formula is parameters × quantization bytes (Q4 ≈ 0.6) + KV cache for your context + ~2GB OS overhead. 16GB is the practical minimum; 32GB is the sweet spot.
Key numbers
- 16GB RAM → up to 13B Q4. 32GB → up to 32B Q4. 64GB → 70B Q4.
- Q4 quantization ≈ 0.6 bytes per parameter; Q8 ≈ 1.07 bytes.
- Add ~2GB OS overhead and 1–8GB KV cache depending on context length.
The weights dominate: an 8B model at Q4 occupies ~4.8GB, a 13B ~7.8GB, a 32B ~19GB, a 70B ~42GB. On top of that, the KV cache grows with context length — a 32B model at 32K context can add 8GB+.
Quantization is the lever: dropping from Q8 to Q4 nearly halves memory use with minimal quality loss. That is why every laptop recommendation we make assumes Q4 as the baseline.
Unified-memory machines (Apple Silicon, Strix Halo) have an edge here because the GPU can use the full system RAM pool — a 64GB MacBook Pro can load a 70B Q4 model that no 32GB Windows laptop can touch.
Related questions
Is 16GB enough for local AI in 2026?
For 7B–13B models, yes. For anything larger or for long-context work, 32GB is the safer buy.
Can I run a 70B model on a laptop?
Only on 64GB+ unified-memory machines (MacBook Pro M4 Max, Strix Halo). Expect 3–8 tokens/sec — usable for batch work, not chat.
Does VRAM vs RAM matter on Windows laptops?
Yes — discrete GPUs use their own VRAM (8–16GB typical), so the model must fit in VRAM for GPU speed. Unified memory avoids that split.
Go deeper
Verified against our benchmark dataset and testing methodology. Facts as of 2026-09-27.