CalcuGrid

Llama 3.2 1B VRAM Requirements

1.24B parameters, 16 layers, 8 key-value heads. At Q4_K_M with a 8K context it needs 1.7 GB — 30 of 30 accelerators here can hold it.

Your setup

About 22.1 GB of that is usable for a model.

The usual choice for running locally. Quality loss is small and it is what most GGUF downloads default to.

8K tokens max 128K

Longer context costs memory linearly — this model uses 32.0 KB per token.

Running Llama 3.2 1B

1.7 GB of memory needed

Runs comfortably on a GeForce RTX 4090

Weights
0.6 GB
KV cache
0.3 GB
Headroom
20.4 GB
Max context
128K
1.7 GB used 22.1 GB usable

Where the memory goes

Llama 3.2 1B memory breakdown
Component Memory
Model weights (Q4_K_M) 0.6 GB
KV cache at 8K tokens 0.3 GB
Runtime allowance 0.8 GB
Total 1.7 GB

Weights are only half the question

A 1.2B model at Q4_K_M is 0.6 GB of weights. That part is fixed — it is the file on disk, and it does not change while the model runs. What changes is the key-value cache, which grows with every token of context you give it.

Llama 3.2 1B spends 32.0 KB per token. At 8K that is 0.3 GB, comfortably smaller than the weights. At its full 128K it is 4.0 GB — larger than the weights themselves. That is the trap: the model loads fine and then fills the card once a conversation gets long.

The reason the cache is as small as it is comes down to grouped-query attention. Llama 3.2 1B has 32 query heads but only 8 key-value heads, and the cache is sized by the latter. Calculators that use the query head count overstate it 4-fold, which is where the wildly pessimistic numbers people quote usually come from. Ties input and output embeddings, which at a 128k vocabulary saves about a fifth of the total parameters.

By quantisation

Llama 3.2 1B memory by quantisation
Format Bits/weight Weights Total at 8K
FP16 / BF16 16 2.3 GB 3.4 GB
Q8_0 8.5 1.2 GB 2.3 GB
Q6_K 6.6 0.9 GB 2.0 GB
Q5_K_M 5.5 0.8 GB 1.8 GB
Q4_K_M 4.5 0.6 GB 1.7 GB
Q3_K_M 3.9 0.6 GB 1.6 GB

By context length

Llama 3.2 1B memory by context length at Q4_K_M
Context KV cache Total
2K tokens 0.1 GB 1.5 GB
4K tokens 0.1 GB 1.6 GB
8K tokens 0.3 GB 1.7 GB
16K tokens 0.5 GB 1.9 GB
32K tokens 1.0 GB 2.4 GB
64K tokens 2.0 GB 3.4 GB
128K tokens 4.0 GB 5.4 GB

Will it run on your hardware?

Which accelerators run Llama 3.2 1B at Q4_K_M
Accelerator Memory Runs it Headroom Max context
GeForce RTX 5090 32 GB Yes 27.7 GB 128K
GeForce RTX 4090 24 GB Yes 20.4 GB 128K
GeForce RTX 3090 24 GB Yes 20.4 GB 128K
GeForce RTX 5080 16 GB Yes 13.0 GB 128K
GeForce RTX 4080 SUPER 16 GB Yes 13.0 GB 128K
GeForce RTX 5070 Ti 16 GB Yes 13.0 GB 128K
GeForce RTX 4070 Ti SUPER 16 GB Yes 13.0 GB 128K
GeForce RTX 5070 12 GB Yes 9.3 GB 128K
GeForce RTX 4070 12 GB Yes 9.3 GB 128K
GeForce RTX 3060 12GB 12 GB Yes 9.3 GB 128K
GeForce RTX 4060 Ti 16GB 16 GB Yes 13.0 GB 128K
GeForce RTX 3080 10 GB Yes 7.5 GB 128K
Radeon RX 7900 XTX 24 GB Yes 20.4 GB 128K
Radeon RX 7900 XT 20 GB Yes 16.7 GB 128K
Radeon RX 9070 XT 16 GB Yes 13.0 GB 128K
2× RTX 3090 (48 GB) 48 GB Yes 42.5 GB 128K
2× RTX 4090 (48 GB) 48 GB Yes 42.5 GB 128K
4× RTX 3090 (96 GB) 96 GB Yes 86.6 GB 128K
Mac (M4 Max, 64 GB) 64 GB Yes 46.3 GB 128K
Mac (M4 Max, 128 GB) 128 GB Yes 94.3 GB 128K
Mac (M4 Pro, 48 GB) 48 GB Yes 34.3 GB 128K
Mac (M4, 24 GB) 24 GB Yes 14.4 GB 128K
Mac (M3 Max, 128 GB) 128 GB Yes 94.3 GB 128K
Mac Studio (M2 Ultra, 192 GB) 192 GB Yes 142.3 GB 128K
Mac (M1 Max, 32 GB) 32 GB Yes 19.7 GB 128K
Mac (M2, 16 GB) 16 GB Yes 9.0 GB 128K
NVIDIA A100 40GB 40 GB Yes 36.3 GB 128K
NVIDIA A100 80GB 80 GB Yes 74.3 GB 128K
NVIDIA H100 80GB 80 GB Yes 74.3 GB 128K
NVIDIA L40S 48GB 48 GB Yes 43.9 GB 128K

At Q4_K_M with a 8K context. "Only just" means the model uses more than 95% of usable memory — it will load and then fall over as soon as anything else touches the device.

Frequently asked questions

How much VRAM does Llama 3.2 1B need?
About 1.7 GB at Q4_K_M with a 8K context — 0.6 GB of weights, 0.3 GB of key-value cache and a 0.8 GB runtime allowance. Unquantised at FP16 it needs 3.4 GB, which is why almost nobody runs it that way locally.
What GPU do I need to run Llama 3.2 1B?
The smallest card here that runs it is the GeForce RTX 3080 with 10 GB. 30 of the 30 accelerators on this page can hold it at the default quantisation. Apple Silicon deserves a mention: unified memory means a Mac can hold models no consumer graphics card can, though it runs them more slowly.
Does context length change how much memory Llama 3.2 1B needs?
Yes, and linearly. This model spends 32.0 KB per token of context, so 8K costs 0.3 GB and its full 128K costs 4.0 GB. On a model with a long context window the cache can end up larger than the quantised weights, which catches people out — the weights fit and then the conversation does not.
Why is the cache smaller than other calculators say?
Because Llama 3.2 1B uses grouped-query attention: 8 key-value heads shared across 32 query heads. The cache is sized by the key-value head count, not the query head count, so using the latter overstates it by a factor of 4. Plenty of calculators still get this wrong.
Which quantisation should I use?
Q4_K_M is the usual answer — the quality loss is small and it is what most GGUF downloads default to. Step up to Q5_K_M or Q6_K if you have memory to spare, since quality improves and the cost is modest. Q8_0 is effectively lossless at half the size of FP16. Drop to Q3_K_M only to fit a larger model that otherwise would not run at all; a bigger model at a rougher quantisation usually beats a smaller one at a fine quantisation.

Architecture: Hugging Face config.json (public mirror of the gated original). GGUF K-quant effective bit widths as produced by llama.cpp.

An estimate of memory, not a benchmark. It counts model weights, the key-value cache at the context you choose, and a fixed runtime allowance. Actual usage moves with the runtime, the batch size, whether flash attention is on, and how much the operating system has already taken from a shared memory pool. Treat a result within a gigabyte of your card's capacity as 'probably not' rather than 'just fits'.

Data on this page last verified .

    ↑↓ to move ↵ to open