Models · chat and writing
Llama 3.2 1B: VRAM requirements and GPUs that run it
Meta's smallest. Fine for simple Q&A and on-device tasks. At Q4 with 8K context it needs about 1.0 GB, so you want a GPU with at least 4 GB of VRAM or a Mac with 8 GB of unified memory.
Memory needed by quantization
| Quant | Weights | Total at 8K | Total at 32K |
|---|---|---|---|
| Q4_K_M | 0.8 GB | 1.0 GB | 1.8 GB |
| Q5_K_M | 0.9 GB | 1.2 GB | 1.9 GB |
| Q8_0 | 1.4 GB | 1.6 GB | 2.4 GB |
| FP16 | 2.6 GB | 2.9 GB | 3.6 GB |
Totals include about 10% runtime overhead and the KV cache. Its conversation memory (KV cache) is unusually small, so going from 8K to 32K context adds only about 0.8 GB. Most people should download Q4_K_M; Q5 or Q8 only make sense if you have spare memory. Read which quantization to pick.
Which GPUs can run Llama 3.2 1B?
40 of the 40 cards we track run it fully in VRAM at 8K context. The cheapest entry point is a 4 GB card such as the GTX 1650, and the fastest is the RTX 5090 at about 920 tokens per second.
| GPU | VRAM | Status | Quant | Speed |
|---|---|---|---|---|
| RTX 5090 | 32 GB | Runs well | Q8_0 | ~920 tok/s |
| RTX 4090 | 24 GB | Runs well | Q8_0 | ~520 tok/s |
| RTX 5080 | 16 GB | Runs well | Q8_0 | ~490 tok/s |
| RX 7900 XTX | 24 GB | Runs well | Q8_0 | ~490 tok/s |
| RTX 3090 / 3090 Ti | 24 GB | Runs well | Q8_0 | ~480 tok/s |
| RTX 5070 Ti | 16 GB | Runs well | Q8_0 | ~460 tok/s |
| RX 7900 XT | 20 GB | Runs well | Q8_0 | ~410 tok/s |
| RTX 3080 | 10 GB | Runs well | Q8_0 | ~390 tok/s |
| RTX 4080 / 4080 Super | 16 GB | Runs well | Q8_0 | ~380 tok/s |
| RTX 5070 | 12 GB | Runs well | Q8_0 | ~340 tok/s |
| RTX 4070 Ti Super | 16 GB | Runs well | Q8_0 | ~340 tok/s |
| RX 9070 XT | 16 GB | Runs well | Q8_0 | ~330 tok/s |
| RX 7800 XT | 16 GB | Runs well | Q8_0 | ~320 tok/s |
| RTX 4090 Laptop | 16 GB | Runs well | Q8_0 | ~290 tok/s |
| Arc A770 | 16 GB | Runs well | Q8_0 | ~290 tok/s |
| Arc A750 | 8 GB | Runs well | Q8_0 | ~260 tok/s |
| RX 6800 XT | 16 GB | Runs well | Q8_0 | ~260 tok/s |
| RTX 4070 / 4070 Super | 12 GB | Runs well | Q8_0 | ~260 tok/s |
| GTX 1080 Ti | 11 GB | Runs well | Q8_0 | ~250 tok/s |
| Arc B580 | 12 GB | Runs well | Q8_0 | ~230 tok/s |
| RTX 5060 Ti | 8 GB | Runs well | Q8_0 | ~230 tok/s |
| RTX 5060 | 8 GB | Runs well | Q8_0 | ~230 tok/s |
| RTX 3070 / 3070 Ti | 8 GB | Runs well | Q8_0 | ~230 tok/s |
| RTX 3060 Ti | 8 GB | Runs well | Q8_0 | ~230 tok/s |
| RTX 5060 Ti | 16 GB | Runs well | Q8_0 | ~230 tok/s |
| RTX 4080 Laptop | 12 GB | Runs well | Q8_0 | ~220 tok/s |
| RX 7700 XT | 12 GB | Runs well | Q8_0 | ~220 tok/s |
| RX 6700 XT | 12 GB | Runs well | Q8_0 | ~200 tok/s |
| RTX 3060 | 12 GB | Runs well | Q8_0 | ~180 tok/s |
| GTX 1660 / 1660 Super | 6 GB | Runs well | Q8_0 | ~170 tok/s |
| RTX 3060 Laptop | 6 GB | Runs well | Q8_0 | ~170 tok/s |
| RTX 4060 Ti | 8 GB | Runs well | Q8_0 | ~150 tok/s |
| RX 7600 | 8 GB | Runs well | Q8_0 | ~150 tok/s |
| RTX 4060 Ti | 16 GB | Runs well | Q8_0 | ~150 tok/s |
| RTX 4060 | 8 GB | Runs well | Q8_0 | ~140 tok/s |
| RTX 4070 Laptop | 8 GB | Runs well | Q8_0 | ~130 tok/s |
| RTX 4060 Laptop | 8 GB | Runs well | Q8_0 | ~130 tok/s |
| RTX 3050 | 8 GB | Runs well | Q8_0 | ~110 tok/s |
| RTX 4050 Laptop | 6 GB | Runs well | Q8_0 | ~98 tok/s |
| GTX 1650 | 4 GB | Runs well | Q8_0 | ~65 tok/s |
Running it on a Mac
Apple Silicon Macs can give a model about 72% of their unified memory. Llama 3.2 1B fits from 8 GB (Apple M1, about 35 tokens per second). With a faster chip like the Apple M3 Ultra (96 GB) expect around 420 tokens per second. Compare the options in our Mac vs NVIDIA guide.
Similar size models: Gemma 3 1B.
Compare Llama 3.2 1B
Questions
How much VRAM does Llama 3.2 1B need?
About 1.0 GB at Q4_K_M with 8K context, 1.6 GB at Q8_0, and 2.9 GB at full FP16. In practice that means a card with at least 4 GB of VRAM for Q4.
Can Llama 3.2 1B run on an 8 GB GPU?
Yes, at Q8_0, using about 1.6 GB.
Can I run Llama 3.2 1B on a Mac?
Yes. The smallest configuration that runs it fully is Apple M1 with 8 GB of unified memory, at about 35 tokens per second. More memory bandwidth (Pro, Max, Ultra chips) means faster replies.
Estimates based on model size, architecture and memory bandwidth. Real numbers vary with software, drivers and context length.