Gemma · Gemma 3 · 1B parameters
Gemma 3 1B VRAM requirements
Gemma 3 1B has 26 layers and uses grouped-query attention (1 KV heads). At Q4_K_M the weights come to 0.6 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.
Runs comfortably
1.4 GB of 21.8 GB · 6%Gemma 3 1B at Q4_K_M leaves 20.4 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Every quantisation of Gemma 3 1B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 1.9 GB | 2.7 GB | Runs comfortably | 32K | 304 |
| FP8 (E4M3) | 8.00 | 0.9 GB | 1.8 GB | Runs comfortably | 32K | 573 |
| INT8 / W8A8 | 8.50 | 0.9 GB | 1.7 GB | Runs comfortably | 32K | 578 |
| Q8_0 (GGUF) | 8.50 | 0.9 GB | 1.7 GB | Runs comfortably | 32K | 578 |
| AWQ 4-bit | 4.25 | 0.9 GB | 1.7 GB | Runs comfortably | 32K | 586 |
| GPTQ 4-bit | 4.25 | 0.9 GB | 1.7 GB | Runs comfortably | 32K | 586 |
| MXFP4 | 4.25 | 0.9 GB | 1.7 GB | Runs comfortably | 32K | 586 |
| Q6_K | 6.56 | 0.8 GB | 1.6 GB | Runs comfortably | 32K | 681 |
| Q5_K_M | 5.67 | 0.7 GB | 1.5 GB | Runs comfortably | 32K | 771 |
| Q5_K_S | 5.52 | 0.6 GB | 1.5 GB | Runs comfortably | 32K | 788 |
| Q4_K_M | 4.85 | 0.6 GB | 1.4 GB | Runs comfortably | 32K | 850 |
| Q4_K_S | 4.58 | 0.6 GB | 1.4 GB | Runs comfortably | 32K | 877 |
| Q4_0 | 4.55 | 0.6 GB | 1.4 GB | Runs comfortably | 32K | 880 |
| IQ4_XS | 4.25 | 0.5 GB | 1.4 GB | Runs comfortably | 32K | 912 |
| Q3_K_M | 3.91 | 0.5 GB | 1.3 GB | Runs comfortably | 32K | 952 |
| IQ3_M | 3.70 | 0.5 GB | 1.3 GB | Runs comfortably | 32K | 978 |
| IQ3_XXS | 3.06 | 0.4 GB | 1.3 GB | Runs comfortably | 32K | 1068 |
| Q2_K | 2.63 | 0.4 GB | 1.2 GB | Runs comfortably | 32K | 1138 |
| IQ2_XXS | 2.06 | 0.4 GB | 1.2 GB | Runs comfortably | 32K | 1247 |
| IQ1_M | 1.75 | 0.3 GB | 1.2 GB | Runs comfortably | 32K | 1315 |
Gemma 3 1B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| H100 SXM 80GB | 80 | 3350 | Runs comfortably | 32K | 2386 |
| A100 80GB | 80 | 2039 | Runs comfortably | 32K | 1563 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 32K | 1477 |
| RTX 5080 | 16 | 960 | Runs comfortably | 32K | 886 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 32K | 850 |
| RTX 5070 Ti | 16 | 896 | Runs comfortably | 32K | 835 |
| RTX 3090 | 24 | 936 | Runs comfortably | 32K | 827 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 32K | 778 |
| L40S | 48 | 864 | Runs comfortably | 32K | 742 |
| RTX A6000 | 48 | 768 | Runs comfortably | 32K | 694 |
| RTX 3080 10GB | 10 | 760 | Runs comfortably | 32K | 688 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 32K | 685 |
| RTX 5070 | 12 | 672 | Runs comfortably | 32K | 647 |
| RTX 4080 Super | 16 | 736 | Runs comfortably | 32K | 643 |
| RTX 4070 Ti Super | 16 | 672 | Runs comfortably | 32K | 593 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 32K | 523 |
| RTX 4070 Super | 12 | 504 | Runs comfortably | 32K | 455 |
| RTX 4070 | 12 | 504 | Runs comfortably | 32K | 455 |
| RTX 5060 Ti 16GB | 16 | 448 | Runs comfortably | 32K | 446 |
| Arc B580 | 12 | 456 | Runs comfortably | 32K | 354 |
| RTX 3060 12GB | 12 | 360 | Runs comfortably | 32K | 345 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 32K | 280 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 32K | 272 |
| RTX 4060 Ti 16GB | 16 | 288 | Runs comfortably | 32K | 268 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 32K | 211 |
Architecture
| Parameters | 1B |
| Layers | 26 |
| Hidden size | 1152 |
| Attention heads / KV heads | 4 / 1 |
| Head dimension | 256 |
| Vocabulary | 262,144 |
| Trained context | 32K |
| Sliding window | 512 (every 6th layer is global) |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | google/gemma-3-1b-it |
The Gemma family
Google's open models. Two things dominate the memory: a 262k vocabulary — on Gemma 3 1B the embedding table is about a third of the file — and sliding-window attention, where only every sixth layer of Gemma 3 sees the full context, and every second layer on Gemma 2. The cache grows far more slowly than the context length suggests.
huggingface.co/google · ai.google.dev/gemma · all 6 Gemma models