32 or 64 GB of RAM for a language model at home
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
RAM (system memory) is what you size when the language model is not fully loaded in VRAM (the graphics card’s memory). That is the usual case on a mini PC with no dedicated GPU, or on a CPU alone. This guide does not replace VRAM and model size. Here you size system memory for Ollama, the browser, and any Docker containers.
Quick read. Find your scenario in the comparison table, then check the machine with When 32 GB is no longer enough before you buy any sticks. Detailed Docker figures sit in the collapsible block at the bottom of the page.
Choose in one step
The decision is one row of the comparison table. Pick the row that looks most like your household (chat only, a Docker stack, a 13B model on the CPU). If the 32 GB column says “at the limit” or “tight,” leave the order alone for now: measure swap during real use first (saturation signs). If saturation is still there after you shrink the model or the services (upgrade order), the 64 GB column is the consistent choice. The questions at the end of the page do not repeat the table. They cover specific cases: DDR5, a mini PC with soldered memory, and an RTX that is already large.
Comparison table
| Scenario | 32 GB | 64 GB |
|---|---|---|
| 8B Q4 chat only | Fine | Extra room for heavy multitasking |
| 13B Q4 on the CPU | At the limit. Close other apps | Fine |
| Docker: Ollama + Open WebUI | Fine | Headroom |
| Docker: plus n8n and Home Assistant | Tight (see the swap signs below) | Recommended, with comfortable headroom |
| Room for the OS and a browser | Moderate | Wide |
The “Docker: plus n8n and Home Assistant” row is spelled out in n8n with a local model and Home Assistant and Ollama.
When 32 GB is no longer enough
Windows (Task Manager → Performance → Memory): find Committed. That is memory Windows and the programs have already reserved, cache included, and it is wider than the “In use” line. When Committed is close to the installed RAM, the bar turns red and the Disk column can climb in percent during an Ollama chat. Windows is swapping (moving memory onto the disk), so replies slow down even when you have no other app open on purpose.
Linux and macOS: the same check, with free -h / htop or Activity Monitor → Swap Used. The details are in the “Task Manager equivalent” question at the end of the page.
What you feel: Ollama replies that get very slow after a few minutes, an interface that freezes, Docker containers killed with an OOM message (out of memory).
A practical rule before you buy sticks: run your usual scenario (the same Ollama model, the same containers) and watch for 5 minutes. If Committed (or swap) stays comfortable, put the upgrade off. If it does not, match the “tight” row in the table with the upgrade order.
RAM vs VRAM (a short reminder)
| Resource | Role |
|---|---|
| VRAM | The model on an NVIDIA or AMD GPU |
| RAM | The model on the CPU, plus the system, plus containers |
On a PC with an RTX card, read VRAM and model size first. RAM still matters for the rest of the system.
A hardware upgrade
If you move to 64 GB, check motherboard compatibility (slot count, DDR4 or DDR5, the manufacturer’s maximum). Indicative stick prices for October 2026: €80–180 for a 32 GB kit, €150–320 for a 64 GB kit. Check the current price on Amazon.com or at your usual retailer before you order.
Checklist before you buy sticks
Measure first. A stick order made with no swap or Committed reading during your real load (Signs) is a guess. Repeat the test after each software step (a smaller model, fewer containers). A hardware upgrade earns its place when the saturation is still there.
- Write down the motherboard or mini PC model and the maximum capacity (from the manufacturer).
- Check DDR4 vs DDR5 and the supported speed.
- Prefer a matched kit of 2 sticks when dual channel is recommended.
- Compare the October 2026 price range across several retailers.
- After you install the sticks, run Ollama again with the same load as before and watch swap (the section “When 32 GB is no longer enough”).
A search for 64 GB DDR5 (October 2026; check motherboard compatibility): 64 GB DDR5 kit on Amazon.com (Amazon affiliate link. Check the current price on Amazon.com. How affiliation works.). To fill out an existing 32 GB kit: 32 GB DDR5 kit on Amazon.com (Amazon affiliate link. Check the current price on Amazon.com.).
Priority: if you are mostly short of disk space for models, upgrade SSD storage for language models before the RAM.
Uses that push toward 64 GB
- n8n and a local model names 32 GB to start, and 64 GB when jobs run in parallel.
- Home Assistant and Ollama
- Ollama on a NAS: the RAM is sometimes fixed, so you choose it when you buy.
Software that goes with it
Once RAM is sized, continue with choose a first model, Docker and Ollama if you containerize, the hardware selector, and the main hardware guide.
Electricity
More RAM does not move idle power in a way you will notice. The CPU or GPU under load is what counts. See the electricity budget.
A worked example (indicative, October 2026)
A 32 GB machine, an 8B Q4 model (about 5 GB), Windows plus a browser (about 6 GB): the “8B Q4 chat only” row stays Fine, with room to spare if you are not measuring swap.
The same machine on the “Docker: plus n8n and Home Assistant” row: run the Committed / swap test again. If saturation remains after you shrink the model, either move to 64 GB or split the work across two hosts (a NAS plus a mini PC).
Dual channel and sticks
On a desktop or a mini PC, two identical sticks (dual channel, two memory channels) sometimes improve throughput. That helps CPU-only language models. Read the motherboard manual (maximum capacity, slot order) before a 2×32 GB kit. LocAIhome does not list motherboard model numbers here.
A mini PC that already has 32 GB
If you are buying a new machine and skipping a RAM upgrade: the 2026 mini PC comparison and noise and power.
Security
More RAM does not raise network risk. Keep Ollama off the public internet. See privacy when AI stays home.
Household examples (October 2026)
A parent, one shared PC, Open WebUI in the evening: 32 GB, one 8B Q4 model, a browser plus Ollama. That is enough when Whisper is not also running on long files (Whisper).
One Docker server: line up with the Docker: plus n8n and Home Assistant row. If swap appears even with sequential n8n workflows, split the services (NAS plus mini PC) before you buy 64 GB as a default.
A gaming tower with 16 GB of RTX VRAM: the model lives mostly in VRAM, and 32 GB of system RAM covers the rest of the PC. 64 GB helps when you keep 50 tabs, a virtual machine, or ComfyUI FLUX open beside the chat. The GPU language model itself is already in VRAM.
Mac mini M4: the “RAM” is Apple Silicon unified memory. Set 24–32 GB in the Apple configurator when you buy. Unified memory is fixed after purchase. The DDR4 and DDR5 stick steps in this guide are for PCs.
A short glossary (Ollama tags)
- instruct: a model variant trained to follow instructions (a dialogue, a summary, a “three bullets” format). Prefer it to the base model for a household. Detail: choose a first model.
- Q4: 4-bit quantization of the GGUF file. The file is lighter, with a bit less nuance than Q8. The Q4_K_M suffix is a common compromise on Ollama in October 2026.
A sensible upgrade order
Before you buy 64 GB “just in case,” try this order:
- A smaller model: move from 13B to an 8B instruct Q4 (choose a first model).
- An NVMe SSD: if disk swap is what dominates (storage for language models).
- Split the services: Ollama on a mini PC, Home Assistant on a NAS, when one box has hit its RAM ceiling.
- 64 GB sticks: when the table above puts your use in “tight” and the swap signs are still there after steps 1–3.
Measure before and after an upgrade
| Step | Windows | Linux / macOS |
|---|---|---|
| Test load | ollama run, plus a browser, plus the usual containers |
The same, plus docker stats |
| What to watch | Task Manager → Memory, Committed | free -h / htop, or Activity Monitor → Swap Used |
| Success | Committed stays under about 85% of installed RAM during the test | Swap stays steady through 5 minutes of chat |
After you install the sticks, run the same scenario again (same model, same containers). Changing the model between the two measurements throws the comparison off.
Docker stack: services, idle RAM, and spikes (rough scale, collapsible)
These October 2026 figures do not replace docker stats on your machine. They explain the Docker rows in the comparison table.
| Service | Role | RAM impact (typical) |
|---|---|---|
| Ollama | The local language-model engine | 4–8 GB, depending on the GGUF model (quantized weights) |
| Open WebUI | A chat interface on the LAN | ~1–2 GB |
| n8n | Automations (mail, PDF) | 1–3 GB, plus spikes when it calls the model |
| Home Assistant | Home automation plus Ollama calls | 2–4 GB, depending on integrations |
| Container / service | Idle RAM, approx. | Spike during a model call |
|---|---|---|
| Ollama alone (8B Q4) | 5–7 GB | +1–2 GB |
| Open WebUI | 1–2 GB | Small |
| n8n | 0.5–1.5 GB | Depends on the workflow |
| Home Assistant | 1–3 GB | Small, unless the integrations are heavy |
A full stack often lands at 12–18 GB before the OS margin. That is why 32 GB is tight when Windows or the browser still uses 6–8 GB.
Software links (after you size the RAM)
Continue with install Ollama, Docker, Ollama, and Open WebUI, and Whisper offline if you add transcription (extra RAM during long jobs). The hardware hub lines up CPU, GPU, and disk.
Common questions
Is 32 GB enough for a family with a single model?
Usually yes, if your row in the comparison table is “8B Q4 chat only” and Committed memory or swap stays comfortable. See the Signs section.
When should I aim for 64 GB?
When your scenario lands on “13B Q4 on the CPU” or “Docker: plus n8n and Home Assistant” in the tight or at-the-limit column, and swap is still there after you shrink the model or the services.
Do I need DDR5?
It is optional. DDR5 helps bandwidth on recent platforms. Check motherboard compatibility before you buy sticks.
My mini PC is soldered at 16 GB. What now?
Skip 13B models. Stay on 3B–7B Q4, or pick another machine in the mini PC comparison. Soldered memory in that case cannot be upgraded.
Does RAM help if I already have a 24 GB RTX?
It does little for the language model alone, because that model lives in VRAM. It helps when you keep many apps, containers, or ComfyUI open.
DDR4 or DDR5 for Ollama on the CPU?
DDR5 can help throughput a little. The gap is often smaller than the choice between 32 and 64 GB, or a fast NVMe SSD.
Where do I see swap on Linux or macOS (the Task Manager equivalent)?
Linux: `free -h` (the Swap line) or `htop` (the swap bar). macOS: Activity Monitor → Memory, the Swap Used figure, and yellow or red memory pressure. Same idea as Windows when the Disk column climbs during an Ollama chat.
Sources
Narrow it down with the hardware selector · French version of this section