How to Deploy Llama 3.2 3B Locally
Llama 3.2 3B is a dense model with 3.2B parameters. At Q4 it needs at least 4GB of VRAM, and 8GB is the comfortable amount for an 8K context. The cheapest card that runs it today is the Tesla V100 16GB at about $305, at roughly 227.3 tokens/s.
Llama 3.2 3B details
How much VRAM does Llama 3.2 3B need for local deployment?
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| 1-bit | 0.8est. | 2.6 | 5.3 | 15.8 |
| 2-bit | 1.2est. | 3.1 | 5.7 | 16.2 |
| 3-bit | 1.9GGUF | 3.4 | 6 | 16.5 |
| 4-bit | 2GGUF | 3.7 | 6.3 | 16.8 |
| 5-bit | 2.3GGUF | 4 | 6.7 | 17.2 |
| 8-bit | 3.4GGUF | 5.2 | 7.8 | 18.3 |
| FP16 | 6.4GGUF | 8.2 | 10.8 | 21.3 |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 128K limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.
Which GPUs can run Llama 3.2 3B?
eBay
| Device | VRAMGB | Price | Est. t/s | Suggested context |
|---|---|---|---|---|
 Tesla V100 16GB | 16 | $30534 listings | 227.3 | 32K |
 Radeon RX 6950 XT | 16 | $5009 listings | 115.3 | 32K |
| GeForce RTX 2080 Ti 22GB | 22 | $5513 listings | 165.1 | 128K |
| Radeon RX 7800 XT | 16 | $56824 listings | 124 | 32K |
| GeForce RTX 5060 Ti 16GB | 16 | $70012 listings | 124.5 | 32K |
| Radeon RX 7900 XT | 20 | $70024 listings | 154.6 | 128K |
| Tesla V100 32GB | 32 | $72023 listings | 227.3 | 128K |
| Instinct MI50 32GB | 32 | $7687 listings | 191.3 | 128K |
| GeForce RTX 5070 | 12 | $77523 listings | 177.9 | 32K |
| Radeon RX 9070 XT | 16 | $79512 listings | 126.8 | 32K |
| Radeon RX 7900 XTX | 24 | $99925 listings | 181.1 | 128K |
| Instinct MI100 | 32 | $9997 listings | 222.7 | 128K |
| GeForce RTX 5070 Ti | 16 | $1,20010 listings | 226.5 | 32K |
| GeForce RTX 4080 | 16 | $1,26330 listings | 188 | 32K |
| GeForce RTX 3090 | 24 | $1,55080 listings | 234.8 | 128K |
| GeForce RTX 5080 | 16 | $1,80014 listings | 239.6 | 32K |
| RTX PRO 4000 Blackwell | 24 | $2,9503 listings | 177.9 | 128K |
| GeForce RTX 4090 | 24 | $3,19940 listings | 249.2 | 128K |
| Radeon PRO W7900 | 48 | $3,4953 listings | 165.3 | 128K |
| RTX PRO 4500 Blackwell | 32 | $4,3003 listings | 226.5 | 128K |
| A100 40GB PCIe | 40 | $4,89913 listings | 346.9 | 128K |
| Instinct MI210 | 64 | $5,04611 listings | 280.3 | 128K |
| MacBook Pro M4 Max 128GB | 128 | $5,5493 listings | 119.8 | 128K |
| GeForce RTX 5090 | 32 | $6,50012 listings | 383.5 | 128K |
| Mac Studio M4 Max 128GB | 128 | $6,5474 listings | 119.8 | 128K |
| MacBook Pro M5 Max 128GB | 128 | $7,2505 listings | 133.1 | 128K |
| RTX PRO 6000 Blackwell | 96 | $16,9854 listings | 383.5 | 128K |
| Mac Studio M3 Ultra 512GB | 512 | $22,00011 listings | 171.4 | 128K |
 Arc Pro B60 | 24 | —No listings | 84.4 | 128K |
| GeForce RTX 5090 D V2 | 24 | —No listings | 311.5 | 128K |
| Arc Pro B65 | 32 | —No listings | 110.1 | 128K |
| Arc Pro B70 | 32 | —No listings | 110.1 | 128K |
| GeForce RTX 4080 SUPER 32GB | 32 | —No listings | 192.4 | 128K |
| GeForce RTX 5090 D | 32 | —No listings | 383.5 | 128K |
| Radeon AI PRO R9700 | 32 | —No listings | 126.8 | 128K |
| Radeon PRO W7800 | 32 | —No listings | 115.3 | 128K |
| CMP 170HX 40GB | 40 | —No listings | 347.7 | 128K |
| GeForce RTX 4090 48GB | 48 | —No listings | 249.2 | 128K |
| RTX PRO 5000 Blackwell 48GB | 48 | —No listings | 311.5 | 128K |
| CMP 170HX 64GB | 64 | —No listings | 336.8 | 128K |
| RTX PRO 5000 Blackwell 72GB | 72 | —No listings | 311.5 | 128K |
| RTX PRO 6000D | 84 | —No listings | 349 | 128K |
| DGX Spark 128GB | 128 | —No listings | 79 | 128K |
| Mac Studio M5 Max 128GB | 128 | —No listings | 133.1 | 128K |
| Ryzen AI Max+ 395 128GB | 128 | —No listings | 54 | 128K |
Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.
Deploy Llama 3.2 3B locally with Ollama, llama.cpp or vLLM
Ollama pulls and runs it in one command; the quantization comes from the official tag.
ollama run llama3.2:3b
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve meta-llama/Llama-3.2-3B-Instruct
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Llama 3.2 3B need?
At Q4_K_M with an 8K context Llama 3.2 3B needs about 4GB of VRAM; 8GB leaves comfortable headroom.
What is the cheapest GPU that runs Llama 3.2 3B?
The Tesla V100 16GB: 16GB of VRAM, currently about $305, at roughly 227.3 tokens/s.
Can you run Llama 3.2 3B with Ollama?
Yes: ollama run llama3.2:3b.