Hardware requirements
What hardware do I need to run Gemma 3 12B?
12.2B
128K context
Gemma Terms of Use
At Q4_K_M the weights are 6.9 GB; with an 8K KV cache and runtime overhead that is 8.3 GB, so it wants a 12 GB card to run with headroom — in practice 12 GB VRAM: GeForce RTX 3060 12 GB and 11 others. Every figure below is computed from the model’s published shape against each device’s usable memory; tokens per second are estimates. The method.
Minimum practical
8 GB VRAM
GeForce RTX 3070 Ti — Q3_K_M, 7.0 GB of 7.0 GB, ~66 tok/s, up to 8K context.
the smallest card that holds it at all, at Q3_K_M — the lowest quantisation worth running
Recommended
12 GB VRAM
GeForce RTX 3060 12 GB — Q4_K_M, 8.3 GB of 10.6 GB, ~32 tok/s, up to 45K context.
runs the recommended quantisation with headroom at 8K context
High quality
20 GB VRAM
Radeon RX 7900 XT — Q8_0, 13.5 GB of 18.4 GB, ~40 tok/s, up to 86K context.
near-lossless Q8_0 with headroom
Long context
24 GB VRAM
GeForce RTX 4090 — Q4_K_M, 15.8 GB of 22.4 GB, ~89 tok/s at 128K.
128K tokens (or the model’s maximum) at the recommended quantisation
Apple Silicon
16 GB unified
M1 · 16 GB — Q4_K_M, 8.3 GB of 10.7 GB, ~5.5 tok/s, up to 46K context.
smallest Mac that runs the recommended quantisation with headroom
CPU only
32 GB RAM
CPU only · DDR4 dual-channel — Q8_0, 13.5 GB of 25.6 GB, ~1.8 tok/s, up to 128K context.
no GPU — weights stream from system RAM
How much VRAM at each quantisation
| Quant | Weights | +KV 8K | +KV 32K | +KV 128K | Total @ 8K | Needs | Quality |
|---|---|---|---|---|---|---|---|
| F16 | 22.7 GB | 0.81 GB | 2.31 GB | 8.3 GB | 24.1 GB | 32 GB card | Reference |
| Q8_0 | 12.1 GB | 0.81 GB | 2.31 GB | 8.3 GB | 13.5 GB | 20 GB card | −0.1% ppl |
| Q6_K | 9.3 GB | 0.81 GB | 2.31 GB | 8.3 GB | 10.7 GB | 16 GB card | −0.4% ppl |
| Q5_K_M | 8.1 GB | 0.81 GB | 2.31 GB | 8.3 GB | 9.5 GB | 16 GB card | −0.8% ppl |
| Q4_K_M | 6.9 GB | 0.81 GB | 2.31 GB | 8.3 GB | 8.3 GB | 12 GB card | −1.9% ppl |
| Q3_K_M | 5.6 GB | 0.81 GB | 2.31 GB | 8.3 GB | 7.0 GB | 10 GB card | −5.4% ppl |
| Q2_K | 4.8 GB | 0.81 GB | 2.31 GB | 8.3 GB | 6.2 GB | 10 GB card | −15% ppl |
Totals add 0.6 GB runtime overhead and an f16 KV cache; "needs" is the smallest discrete card class whose usable memory (after display and driver reserve) leaves 15% headroom. A q8_0 cache roughly halves the KV columns.
NVIDIA · GeForce
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| GeForce RTX 5090 | 32 GB | Q8_0 | Q8_0 | 128K | 158 | yes | Check |
| GeForce RTX 5080 | 16 GB | Q6_K | Q8_0 | 106K | 85 | no | Check |
| GeForce RTX 5070 Ti | 16 GB | Q6_K | Q8_0 | 106K | 79 | no | Check |
| GeForce RTX 5070 | 12 GB | Q4_K_M | Q5_K_M | 45K | 59 | no | Check |
| GeForce RTX 5060 Ti 16 GB | 16 GB | Q6_K | Q8_0 | 106K | 40 | no | Check |
| GeForce RTX 5060 | 8 GB | — | Q3_K_M | — | — | no | Check |
| GeForce RTX 4090 | 24 GB | Q8_0 | Q8_0 | 128K | 89 | yes | Check |
| GeForce RTX 4080 Super | 16 GB | Q6_K | Q8_0 | 106K | 65 | no | Check |
| GeForce RTX 4080 | 16 GB | Q6_K | Q8_0 | 106K | 63 | no | Check |
| GeForce RTX 4070 Ti Super | 16 GB | Q6_K | Q8_0 | 106K | 59 | no | Check |
| GeForce RTX 4070 Ti | 12 GB | Q4_K_M | Q5_K_M | 45K | 44 | no | Check |
| GeForce RTX 4070 Super | 12 GB | Q4_K_M | Q5_K_M | 45K | 44 | no | Check |
| GeForce RTX 4070 | 12 GB | Q4_K_M | Q5_K_M | 45K | 44 | no | Check |
| GeForce RTX 4060 Ti 16 GB | 16 GB | Q6_K | Q8_0 | 106K | 25 | no | Check |
| GeForce RTX 4060 Ti | 8 GB | — | Q3_K_M | — | — | no | Check |
| GeForce RTX 4060 | 8 GB | — | Q3_K_M | — | — | no | Check |
| GeForce RTX 3090 Ti | 24 GB | Q8_0 | Q8_0 | 128K | 89 | yes | Check |
| GeForce RTX 3090 | 24 GB | Q8_0 | Q8_0 | 128K | 83 | yes | Check |
| GeForce RTX 3080 Ti | 12 GB | Q4_K_M | Q5_K_M | 45K | 80 | no | Check |
| GeForce RTX 3080 12 GB | 12 GB | Q4_K_M | Q5_K_M | 45K | 80 | no | Check |
| GeForce RTX 3080 | 10 GB | Q3_K_M | Q4_K_M | 16K | 67 | no | Check |
| GeForce RTX 3070 Ti | 8 GB | — | Q3_K_M | — | — | no | Check |
| GeForce RTX 3070 | 8 GB | — | Q3_K_M | — | — | no | Check |
| GeForce RTX 3060 Ti | 8 GB | — | Q3_K_M | — | — | no | Check |
| GeForce RTX 3060 12 GB | 12 GB | Q4_K_M | Q5_K_M | 45K | 32 | no | Check |
| GeForce RTX 2080 Ti | 11 GB | Q3_K_M | Q5_K_M | 30K | 54 | no | Check |
| GeForce RTX 2060 12 GB | 12 GB | Q4_K_M | Q5_K_M | 45K | 30 | no | Check |
| GeForce GTX 1080 Ti | 11 GB | Q3_K_M | Q5_K_M | 30K | 43 | no | Check |
NVIDIA · GeForce laptop
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| GeForce RTX 5090 Laptop | 24 GB | Q8_0 | Q8_0 | 128K | 79 | yes | Check |
| GeForce RTX 4090 Laptop | 16 GB | Q6_K | Q8_0 | 106K | 51 | no | Check |
| GeForce RTX 4080 Laptop | 12 GB | Q4_K_M | Q5_K_M | 45K | 38 | no | Check |
| GeForce RTX 4070 Laptop | 8 GB | — | Q3_K_M | — | — | no | Check |
NVIDIA · Workstation
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| RTX 6000 Ada | 48 GB | Q8_0 | Q8_0 | 128K | 85 | yes | Check |
| RTX 5000 Ada | 32 GB | Q8_0 | Q8_0 | 128K | 51 | yes | Check |
| RTX 4000 Ada | 20 GB | Q8_0 | Q8_0 | 128K | 32 | tight | Check |
| RTX 2000 Ada | 16 GB | Q6_K | Q8_0 | 106K | 20 | no | Check |
| RTX A6000 | 48 GB | Q8_0 | Q8_0 | 128K | 68 | yes | Check |
| RTX A5000 | 24 GB | Q8_0 | Q8_0 | 128K | 68 | yes | Check |
| RTX A4000 | 16 GB | Q6_K | Q8_0 | 106K | 40 | no | Check |
NVIDIA · Data centre
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| L40S | 48 GB | Q8_0 | Q8_0 | 128K | 76 | yes | Check |
| L4 | 24 GB | Q8_0 | Q8_0 | 128K | 26 | yes | Check |
| A10 | 24 GB | Q8_0 | Q8_0 | 128K | 53 | yes | Check |
| A100 40 GB | 40 GB | Q8_0 | Q8_0 | 128K | 137 | yes | Check |
| A100 80 GB | 80 GB | Q8_0 | Q8_0 | 128K | 180 | yes | Check |
| H100 PCIe | 80 GB | Q8_0 | Q8_0 | 128K | 176 | yes | Check |
| H100 SXM | 80 GB | Q8_0 | Q8_0 | 128K | 296 | yes | Check |
| H200 SXM | 141 GB | Q8_0 | Q8_0 | 128K | 424 | yes | Check |
| Tesla P40 | 24 GB | Q8_0 | Q8_0 | 128K | 31 | yes | Check |
AMD · Radeon
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| Radeon RX 9070 XT | 16 GB | Q6_K | Q8_0 | 106K | 57 | no | Check |
| Radeon RX 9070 | 16 GB | Q6_K | Q8_0 | 106K | 57 | no | Check |
| Radeon RX 7900 XTX | 24 GB | Q8_0 | Q8_0 | 128K | 85 | yes | Check |
| Radeon RX 7900 XT | 20 GB | Q8_0 | Q8_0 | 128K | 71 | tight | Check |
| Radeon RX 7900 GRE | 16 GB | Q6_K | Q8_0 | 106K | 51 | no | Check |
| Radeon RX 7800 XT | 16 GB | Q6_K | Q8_0 | 106K | 55 | no | Check |
| Radeon RX 7700 XT | 12 GB | Q4_K_M | Q5_K_M | 45K | 38 | no | Check |
| Radeon RX 7600 XT | 16 GB | Q6_K | Q8_0 | 106K | 25 | no | Check |
| Radeon RX 6900 XT | 16 GB | Q6_K | Q8_0 | 106K | 45 | no | Check |
| Radeon RX 6800 XT | 16 GB | Q6_K | Q8_0 | 106K | 45 | no | Check |
| Radeon RX 6700 XT | 12 GB | Q4_K_M | Q5_K_M | 45K | 34 | no | Check |
AMD · Radeon Pro
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| Radeon Pro W7900 | 48 GB | Q8_0 | Q8_0 | 128K | 76 | yes | Check |
| Radeon Pro W7800 | 32 GB | Q8_0 | Q8_0 | 128K | 51 | yes | Check |
AMD · Instinct
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| Instinct MI210 | 64 GB | Q8_0 | Q8_0 | 128K | 145 | yes | Check |
| Instinct MI300X | 192 GB | Q8_0 | Q8_0 | 128K | 468 | yes | Check |
AMD · Ryzen AI Max
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| Ryzen AI Max+ 395 · 128 GB | 128 GB | Q8_0 | Q8_0 | 128K | 21 | yes | Check |
| Ryzen AI Max+ 395 · 64 GB | 64 GB | Q8_0 | Q8_0 | 128K | 21 | yes | Check |
Apple · Apple Silicon
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| M1 · 16 GB | 16 GB | Q4_K_M | Q5_K_M | 46K | 5.5 | no | Check |
| M1 Pro · 32 GB | 32 GB | Q8_0 | Q8_0 | 128K | 16 | yes | Check |
| M1 Max · 64 GB | 64 GB | Q8_0 | Q8_0 | 128K | 33 | yes | Check |
| M1 Ultra · 128 GB | 128 GB | Q8_0 | Q8_0 | 128K | 65 | yes | Check |
| M2 · 24 GB | 24 GB | Q8_0 | Q8_0 | 128K | 8.1 | tight | Check |
| M2 Pro · 32 GB | 32 GB | Q8_0 | Q8_0 | 128K | 16 | yes | Check |
| M2 Max · 96 GB | 96 GB | Q8_0 | Q8_0 | 128K | 33 | yes | Check |
| M2 Ultra · 192 GB | 192 GB | Q8_0 | Q8_0 | 128K | 65 | yes | Check |
| M3 · 24 GB | 24 GB | Q8_0 | Q8_0 | 128K | 8.1 | tight | Check |
| M3 Pro · 36 GB | 36 GB | Q8_0 | Q8_0 | 128K | 12 | yes | Check |
| M3 Max · 64 GB | 64 GB | Q8_0 | Q8_0 | 128K | 33 | yes | Check |
| M3 Max · 128 GB | 128 GB | Q8_0 | Q8_0 | 128K | 33 | yes | Check |
| M3 Ultra · 256 GB | 256 GB | Q8_0 | Q8_0 | 128K | 65 | yes | Check |
| M3 Ultra · 512 GB | 512 GB | Q8_0 | Q8_0 | 128K | 65 | yes | Check |
| M4 · 32 GB | 32 GB | Q8_0 | Q8_0 | 128K | 9.8 | yes | Check |
| M4 Pro · 48 GB | 48 GB | Q8_0 | Q8_0 | 128K | 22 | yes | Check |
| M4 Pro · 64 GB | 64 GB | Q8_0 | Q8_0 | 128K | 22 | yes | Check |
| M4 Max · 64 GB | 64 GB | Q8_0 | Q8_0 | 128K | 33 | yes | Check |
| M4 Max · 128 GB | 128 GB | Q8_0 | Q8_0 | 128K | 44 | yes | Check |
Intel · Arc
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| Arc B580 | 12 GB | Q4_K_M | Q5_K_M | 45K | 40 | no | Check |
| Arc B570 | 10 GB | Q3_K_M | Q4_K_M | 16K | 34 | no | Check |
| Arc A770 16 GB | 16 GB | Q6_K | Q8_0 | 106K | 49 | no | Check |
| Arc A750 | 8 GB | — | Q3_K_M | — | — | no | Check |
CPU · CPU + system RAM
| Device | Memory | Best quant with headroom | Fits at all | Max context @ Q4_K_M | Tok/s est. | 128K | |
|---|---|---|---|---|---|---|---|
| CPU only · DDR4 dual-channel | 32 GB | Q8_0 | Q8_0 | 128K | 3.1 | yes | Check |
| CPU only · DDR5 dual-channel | 64 GB | Q8_0 | Q8_0 | 128K | 5.5 | yes | Check |
| CPU only · DDR5 8-channel server | 256 GB | Q8_0 | Q8_0 | 128K | 19 | yes | Check |