Llama 3.2 3B requirements: measured RAM, speed and the server you need

Llama 3.2 3B requirements are modest: in Ollama it needs about 3 GiB of system RAM and no GPU. With the model loaded, our five CPU-only servers had 3002 to 3282 MiB of RAM in use, operating system included. A 4 GB VPS runs it (5.8 tokens/s on a 2 vCPU Hetzner CX23); a 4 vCPU, 8 GB server is the comfortable choice (13.5 tokens/s on the CX33).

Why this matters now: 3B models are now good enough for summaries, classification and simple agents, and they are the largest size that fits on the cheapest VPS plans. Published requirement figures for this model range from 1.8 GB to 6 GB because they are estimates. These are measurements. More models and servers are in our local AI section.

Tested in the Lab Hetzner CX23, CX33, CX43, CAX11, CAX21 · Nov 2026

What the official sources say

Meta's model card on Hugging Face lists Llama 3.2 3B at 3.21B parameters (billion weights), with a 128K-token context, released on 25 September 2024, pretrained on up to 9 trillion tokens with a December 2023 knowledge cutoff. It officially supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai, under the Llama 3.2 Community License.

The card does not state hardware requirements for servers. Its only device figures are from a OnePlus 12 phone, using Meta's own 4-bit formats (SpinQuant, QLoRA) on the ExecuTorch runtime: for example about 3,726 MB of memory and 19.7 tokens/s with SpinQuant. That is a different runtime and quantization from Ollama, so do not use it to size a VPS.

The Ollama library lists llama3.2:3b (also the default llama3.2 tag) as a 2.0 GB download, Q4_K_M quantization, 3.21B parameters, 128K context. Q4_K_M means the weights are stored in roughly 4 to 5 bits each instead of 16. The Ollama FAQ says it uses a 4096-token context by default, which is what we tested.

Measured: RAM, speed and load time on five servers

We ran llama3.2:3b with Ollama 0.40.14, CPU-only, on five Hetzner Cloud servers in Helsinki on 5 November 2026. "Generation tokens/s" is how fast the answer is written (a token is a word or a piece of a word). "Prompt tokens/s" is how fast the input is read. "RAM used" is whole-system memory in use right after loading, from free -m, in MiB (1 MiB = 1,048,576 bytes, about 5 % more than a MB).

Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data

RAM: what is left on a 4 GB server

The "used" figure includes more than the model. A fresh server used 308 to 500 MiB before we installed anything, and an idle Docker daemon from an earlier step of our test was still running. On your server, with no Docker, it will be a little lower; with n8n, a database or a web UI beside it, much higher. That is why a 4 GB server is the minimum, not the recommendation: 591 MiB left on the CAX11 is not enough for n8n next to the model: on our own n8n server, n8n alone used 785.3 MiB right after start. Add up your own stack with the RAM calculator:

Speed: what each server feels like

A 256-token answer (a few paragraphs) takes about 44 seconds at 5.8 tokens/s on the CX23, 39 seconds at 6.5 on the CAX11, 21 seconds at 12.2 on the CAX21, 19 seconds at 13.5 on the CX33 and 12 seconds at 21.2 on the CX43. Reading the question is much faster: 325.2 tokens/s on the CX23 up to 1161.3 tokens/s on the CX43, so the first words appear quickly on all five.

The first request after an idle period also pays the load time: 3.7 s on the CX33 and CX43, 5.2 s on the CAX21, 6 s on the CAX11 and 8.8 s on the CX23. Ollama keeps the model loaded for 5 minutes after the last request by default, per its FAQ.

Minimum and comfortable server for Llama 3.2 3B

Prices are from the Hetzner Cloud API, gross, checked on 5 November 2026. The ARM CAX21 (12.2 tokens/s, 12.69 €) costs more than the CX33 and is a little slower on this model, so we do not list it as a pick. Our VPS for Ollama guide compares all five per model size, and the full CPU benchmark explains why vCPU count helped more than single-core speed.

3B or 1B on the same server?

If a 4 GB server feels too tight, the smaller Llama 3.2 1B is the alternative. On the same five servers it used 2001 to 2315 MiB of system RAM, 863 to 1011 MiB less than 3B on each server, and generated 1.7 to 2.3 times as fast: 10.3 tokens/s on the CX23 against 5.8 for 3B, 15 against 6.5 on the CAX11, 23.2 against 13.5 on the CX33. The 1B download is 1.3 GB in the Ollama library.

What you give up is quality. We have not scored answer quality in the Lab, so we cannot put a number on it. Our practical rule: use 1B for short, structured jobs (sort this email, pull out the date, answer yes or no), and 3B when the output is text a person reads. If 3B fits your RAM with room to spare, there is little reason to drop to 1B for speed alone on a 4 vCPU server.

Check it on your own server

After pulling the model and sending one request, run ollama ps. The Ollama FAQ explains its Processor column: "100% CPU" means the model is loaded entirely in system memory, which is what you will see on any VPS without a GPU. Then compare free -m with our table above. If your "used" figure is far above 3300 MiB with only Ollama running, something else on the server is taking memory.

What changes the requirements

• Longer context. The model supports 128K tokens, but every token of context needs memory. Our numbers are for Ollama's 4096-token default. We have not measured RAM at longer contexts yet, so if you plan to feed it long documents, leave more headroom than the table shows.
• Other quantizations. Higher-precision versions (for example Q8 or full 16-bit) are larger downloads and need more RAM. We tested only the default Q4_K_M.
• Several users. Ollama processes one request per model at a time by default (OLLAMA_NUM_PARALLEL is 1). Our speeds are for a single request.
• A GPU. Not needed for this model size. If you have bigger plans, see CPU vs GPU for local LLMs.

When 3B is not enough, the next step up is 8B, which needs roughly twice the RAM: it used 5796 to 6023 MiB on our 8 GB and 16 GB servers and ran out of memory on both 4 GB servers. Details in Llama 3.1 8B requirements.

Next step: run it yourself with our guide to install Ollama on a VPS, then ollama pull llama3.2:3b.