Llama 3.2 1B & 3B: Ollama Tags, RAM and Speed

Llama 3.2 3B Ollama tags and sizes by quantisation

llama3.2:3b, llama3.2 and llama3.2:latest are the same file: the Q4_K_M build of the Instruct model, 2.02 GB. The other quantisations are separate tags. Sizes are from the Ollama registry; the memory column is arithmetic (file + key-value cache at 4K + about 1 GB of buffers), not a measurement.

Base (non-instruct) builds exist too, such as llama3.2:3b-text-q4_K_M (2.02 GB), for completion and fine-tuning work. See quantization explained for what each level trades away, and the Ollama RAM and VRAM table for other models.

How much RAM do you need?

The model needs its file plus a small cache. Llama 3.2 uses grouped-query attention (8 key-value heads), so the 3B's cache is only about 115 KB per token: 0.47 GB at 4K, 3.8 GB at 32K and 15 GB at the full 128K. What else is running matters as much as the model.

• 4 GB system RAM: the 1B at Q4_K_M (about 1.9 GB in use). The 3B at Q3_K_M (about 3.2 GB) will swap with an operating system loaded.
• 4 GB graphics card: the 3B at Q4_K_M fits entirely in VRAM at 4K context.
• 8 GB laptop or Mac: the 3B at any quantisation up to Q8_0, with a browser open. Raise the context to 16-32K if you need it.
• 16 GB: the 3B at FP16, or Q4_K_M with a 64K+ context. At this size an 8B model such as Llama 3.1 8B also fits and is noticeably stronger.
• Raspberry Pi 5 (8 GB): runs the 3B on CPU; see the speed section for what its memory bandwidth allows.

For a ranked list by memory size, see the best local AI models for 8 GB RAM and the small language models guide.

How fast does Llama 3.2 3B run?

Generation speed is capped by how fast the hardware can read the weights: memory bandwidth divided by file size. That ceiling is an upper bound, not a prediction; real output lands below it.

Bandwidth from Apple's and NVIDIA's spec pages; the DDR5 and Raspberry Pi figures are transfer rate times bus width. CPU-only machines usually land well under their ceiling because the CPU also has to do the arithmetic. Background: why memory bandwidth sets LLM speed and running an LLM on CPU only.

Llama 3.2 1B

The 1B is the smallest Llama: 1.23B parameters, 16 layers, the same 128K context. Note that llama3.2:1b pulls the Q8_0 build (1.32 GB), not Q4_K_M; ask for llama3.2:1b-instruct-q4_K_M (0.81 GB) when every megabyte counts. Its cache is about 33 KB per token, so even the full 128K context adds only 4.3 GB.

The trade-off is real: Meta's card gives the 1B 49.3 on MMLU and 25.7 on tool calling, against 63.4 and 67.0 for the 3B. It is fine for classification, short extraction, autocomplete of plain text and as a fast draft model; for anything that needs reasoning or reliable tool calls, use the 3B.

Published benchmarks

Meta's own figures for the instruction-tuned models (bf16), from the Llama 3.2 model card. This site has not re-run them; a 4-bit download will score slightly lower.

What Llama 3.2 3B is good for

• Summarising and rewriting. Meta positioned the small models for summarisation, instruction following and rewriting on-device; the 77.4 IFEval score backs the instruction-following part.
• Tool calling and simple agents. 67.0 on BFCL V2 is high for its size, so it can route a request to a function with structured arguments. See the best Ollama models for agents.
• Private, offline helpers. Note-taking assistants, local email triage, a home-automation brain on a mini PC: jobs where the data should not leave the machine and a 2 GB model is fast enough.
• Draft model. As the small model in speculative decoding or a two-stage pipeline in front of a larger Llama.
• Not for: long code generation, multi-step maths or anything where a wrong answer is costly. An 8B model is the next step up.

Phones and edge devices

Meta announced the 1B and 3B as its first Llama models built for edge and mobile devices, and later published its own quantised builds (SpinQuant and QLoRA), which the model card notes run with an 8K context. Ollama itself targets macOS, Windows and Linux, so on a phone you would use an app that bundles llama.cpp or MLC instead; the 1B at 4-bit is the comfortable size for phones with 6-8 GB.

On single-board computers the binding limit is memory bandwidth, not compute: the Raspberry Pi 5's ceiling for the 3B is about 8 tok/s, so the 1B is the better fit there.

Install and run

# Install Ollama (Linux; macOS and Windows use the installer from ollama.com)
curl -fsSL https://ollama.com/install.sh | sh

# Llama 3.2 3B, Q4_K_M, 2.02 GB
ollama run llama3.2:3b

# Llama 3.2 1B, Q8_0, 1.32 GB
ollama run llama3.2:1b

# Smaller 3B for a 4 GB machine, 1.69 GB
ollama pull llama3.2:3b-instruct-q3_K_M

To use more than the default 4K context, type /set parameter num_ctx 32768 inside ollama run, or pass num_ctx in the API options. On Windows, follow the Ollama Windows installation guide.

Alternatives to Llama 3.2 3B

Other small models in the same memory class. MMLU is each vendor's own figure, so treat small gaps as noise.

Frequently asked questions