# Gemma 3 12B: hardware requirements
> Gemma 3 12B (Google) is a 12.2B dense model with a 128K-token context window, released under the Gemma Terms of Use. It needs about 9GB of memory at Q4_K_M with 8K context.
Gemma 3 12B was Google's mid-size open model for 2025, with image input, support for over 140 languages, and a 128K context window. Gemma 4 has replaced it: Gemma 4 E4B covers the smaller tier and Gemma 4 26B A4B is far stronger for modestly more memory, both under the more permissive Apache 2.0 license. At Q4_K_M it needs about 9GB with 8K context, so it fits 12GB GPUs comfortably and 16GB cards at Q6_K or Q8_0. Its interleaved sliding-window attention keeps long-context memory low.

## Recommended hardware
- **Minimum GPU:** RTX 5060 (8GB), Q3_K_M, ~58 tokens/sec
- **Best value:** RTX 3060 12GB (12GB), Q5_K_M, ~29 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q8_0, ~105 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~15 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 4.5GB | 6.1GB | 7.7GB |
| Q3_K_M | 5.9GB | 7.5GB | 9.1GB |
| Q4_K_M | 7.3GB | 9.0GB | 10.6GB |
| Q5_K_M | 8.6GB | 10.3GB | 11.9GB |
| Q6_K | 9.9GB | 11.7GB | 13.3GB |
| Q8_0 | 12.5GB | 14.3GB | 15.9GB |
| BF16 | 23.5GB | 25.7GB | 27.3GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q8_0 | ~105 tok/s | ~$4,300 |
| RTX PRO 6000 | 96GB | Q8_0 | ~105 tok/s | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~79 tok/s | ~$9,100 |
| RTX 5080 | 16GB | Q6_K | ~72 tok/s | ~$1,600 |
| RTX 5070 Ti | 16GB | Q6_K | ~67 tok/s | ~$1,200 |
| RTX 4090 | 24GB | Q8_0 | ~61 tok/s | ~$2,500 used |
| RTX 5070 | 12GB | Q5_K_M | ~59 tok/s | ~$800 |
| RTX 6000 Ada | 48GB | Q8_0 | ~58 tok/s | ~$4,750 used |
| RTX 5060 | 8GB | Q3_K_M | ~58 tok/s | ~$450 |
| RTX 4080 Super | 16GB | Q6_K | ~56 tok/s | ~$1,150 used |
| RTX 4070 Ti Super | 16GB | Q6_K | ~51 tok/s | ~$900 used |
| RTX 3090 | 24GB | Q8_0 | ~51 tok/s | ~$1,050 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~50 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~50 tok/s | ~$9,499 |
| RX 7900 XTX | 24GB | Q8_0 | ~46 tok/s | ~$1,050 used |
| RTX 4070 | 12GB | Q5_K_M | ~45 tok/s | ~$560 used |
| V100 32GB | 32GB | Q8_0 | ~44 tok/s | ~$700 used |
| RTX A6000 | 48GB | Q8_0 | ~42 tok/s | ~$4,400 used |
| RX 9070 XT | 16GB | Q6_K | ~39 tok/s | ~$800 |
| RTX 5060 Ti 16GB | 16GB | Q6_K | ~35 tok/s | ~$620 |
| MI50 32GB | 32GB | Q8_0 | ~34 tok/s | ~$600 used |
| M3 Ultra 512GB | 512GB | Q8_0 | ~34 tok/s | $9,499 MSRP |
| R9700 | 32GB | Q8_0 | ~32 tok/s | ~$1,800 |
| M5 Max 128GB | 128GB | Q8_0 | ~30 tok/s | ~$5,100 |
| RTX 3060 12GB | 12GB | Q5_K_M | ~29 tok/s | ~$270 used |

## How to run it
- Ollama: `ollama run gemma3:12b`
- llama.cpp: `llama-server -hf bartowski/google_gemma-3-12b-it-GGUF:Q4_K_M -c 8192 -ngl 99`
- MLX (Mac): `mlx_lm.chat --model mlx-community/gemma-3-12b-it-4bit`
- Weights: https://huggingface.co/google/gemma-3-12b-it

## Strengths
- Good conversational quality and writing for its size
- Image input included
- Sliding-window attention keeps the KV cache small at long context

## Weaknesses
- Superseded by Gemma 4, which is stronger at every size
- No reasoning mode; weak at math and multi-step problems
- Gemma license is more restrictive than the Apache 2.0 license used by Gemma 4
Full page: https://llamabuilds.ai/models/gemma-3-12b · Hardware guide: https://llamabuilds.ai/models/gemma-3-12b/requirements

---
Source: https://llamabuilds.ai/models/gemma-3-12b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.