What GPU do I need to run meta-llama/Llama-3.2-3B?
3.2B parameters, published in BF16. View on Hugging FaceGated
Set up Llama-3.2-3BLlama-3.2-3B is published by meta-llama on Hugging Face, with 378,160 downloads and 902 likes to date. It's a LlamaForCausalLM model built for text-generation, published natively in BF16, and gated: you'll need to accept the model's terms on Hugging Face before downloading weights.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| BF16 | 6.0 GB | 7.2 GB | RTX 3060 | 1 | $0.110/hr |
| FP8 (quantized) | 3.0 GB | 3.6 GB | RTX 5060 Ti | 1 | $0.160/hr |
| INT4 (quantized) | 1.5 GB | 1.8 GB | RTX 3060 | 1 | $0.110/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run Llama-3.2-3B at its published (BF16) precision: 1× RTX 3060, at $0.110/hr per GPU ($0.110/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
Llama-3.2-3B: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Grouped-query attention: every layer caches keys and values for a few shared KV heads. Cached per token: 114,688 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.44 GB |
| 32K | 3.5 GB |
| 128K | 14.0 GB |
Computed from the layer, head and window counts in the model's own configuration (unsloth/Llama-3.2-3B), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
Llama-3.2-3B: common questions
Does Llama-3.2-3B fit on a 8 GB GPU?
Yes. At BF16 it needs 7.2 GB of VRAM, so an 8 GB card holds it with 0.8 GB to spare. A 6 GB card is not enough for it at BF16.
Do I need approval to download Llama-3.2-3B?
Yes. meta-llama gates this repository on Hugging Face, so you have to accept its terms with a signed-in account before the weights will download. It does not change the 7.2 GB the model needs once you have them.
What is the least VRAM Llama-3.2-3B can run in?
1.8 GB, at INT4 (quantized), which fits a 6 GB card, against 7.2 GB at BF16. That row assumes an INT4 checkpoint has actually been published for this model, so check its Hugging Face page before planning around it.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Llama 3.2 models
- Llama-3.2-3B-Instruct (3.2B, BF16)
- Llama-3.2-3B-Instruct-pythonic (3.2B, BF16)
- NetrAI-L3 (3.2B, F16)
- alloma-3B-Instruct (3.2B, BF16)
- svara-tts-v1 (3.3B, F32)
Lineage
Also published as: unsloth/Llama-3.2-3B. Identical copies of this model under another account.
Alternatives at this size
Other models for text-generation within about a third of Llama-3.2-3B's 3.2B parameters, from other model lines.
- Qwen2.5-3B-Instruct (3.1B, BF16)
- phi-2 (2.8B, F16)
- PowerMoE-3b (3.4B, F32)
- gemma-2-2b-it (2.6B, BF16)
- SmolLM3-3B-Base (3.1B, BF16)
More on Llama-3.2-3B
Fits on an 8 GB GPU at BF16: every model that fits in 8 GB.
Best chat and assistants models: how Llama-3.2-3B ranks against the rest.
Related reading: H100 pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
