# Gemma 3 27B: hardware requirements
> Gemma 3 27B (Google) is a 27.4B dense model with a 128K-token context window, released under the Gemma Terms of Use. It needs about 19GB of memory at Q4_K_M with 8K context.
Gemma 3 27B was one of the most popular 24GB-GPU models of 2025, praised for natural writing, image understanding, and broad multilingual support. Gemma 4 31B replaces it with a reasoning mode, a 256K context, much stronger coding and math, and an Apache 2.0 license, so it is now superseded. At Q4_K_M it needs about 18GB with 8K context, leaving room for long context on a 24GB card, and Google's quantization-aware Q4_0 release keeps quality close to the original at a similar size. Its sliding-window design keeps the KV cache small, about 0.7GB per 8K tokens.

## Recommended hardware
- **Minimum GPU:** RTX 4060 Ti 16GB (16GB), Q3_K_M, ~17 tokens/sec
- **Best value:** RTX 3090 (24GB), Q4_K_M, ~39 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q6_K, ~60 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~6.6 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 10.2GB | 12.2GB | 14.2GB |
| Q3_K_M | 13.3GB | 15.4GB | 17.4GB |
| Q4_K_M | 16.6GB | 18.7GB | 20.7GB |
| Q5_K_M | 19.5GB | 21.7GB | 23.7GB |
| Q6_K | 22.5GB | 24.9GB | 26.9GB |
| Q8_0 | 28.7GB | 31.2GB | 33.3GB |
| BF16 | 54.0GB | 57.3GB | 59.3GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q6_K | ~60 tok/s | ~$4,300 |
| RTX 5080 | 16GB | Q3_K_M | ~55 tok/s | ~$1,600 |
| RTX 5070 Ti | 16GB | Q3_K_M | ~51 tok/s | ~$1,200 |
| RTX PRO 6000 | 96GB | Q8_0 | ~48 tok/s | ~$16,000 |
| RTX 4090 | 24GB | Q4_K_M | ~46 tok/s | ~$2,500 used |
| RTX 4080 Super | 16GB | Q3_K_M | ~42 tok/s | ~$1,150 used |
| RTX 3090 | 24GB | Q4_K_M | ~39 tok/s | ~$1,050 used |
| RTX 4070 Ti Super | 16GB | Q3_K_M | ~39 tok/s | ~$900 used |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~36 tok/s | ~$9,100 |
| RX 7900 XTX | 24GB | Q4_K_M | ~36 tok/s | ~$1,050 used |
| RX 9070 XT | 16GB | Q3_K_M | ~30 tok/s | ~$800 |
| RTX 5060 Ti 16GB | 16GB | Q3_K_M | ~26 tok/s | ~$620 |
| RTX 6000 Ada | 48GB | Q8_0 | ~26 tok/s | ~$4,750 used |
| V100 32GB | 32GB | Q6_K | ~25 tok/s | ~$700 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~23 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~23 tok/s | ~$9,499 |
| MI50 32GB | 32GB | Q6_K | ~20 tok/s | ~$600 used |
| RTX A6000 | 48GB | Q8_0 | ~19 tok/s | ~$4,400 used |
| R9700 | 32GB | Q6_K | ~18 tok/s | ~$1,800 |
| RTX 4000 Ada | 20GB | Q4_K_M | ~17 tok/s | $1,250 MSRP |
| RTX 4060 Ti 16GB | 16GB | Q3_K_M | ~17 tok/s | ~$380 used |
| M3 Ultra 512GB | 512GB | Q8_0 | ~15 tok/s | $9,499 MSRP |
| RX 9060 XT 16GB | 16GB | Q3_K_M | ~15 tok/s | ~$470 |
| Arc Pro B60 | 24GB | Q4_K_M | ~15 tok/s | ~$650 |
| Arc Pro B70 | 32GB | Q6_K | ~14 tok/s | ~$1,300 |

## How to run it
- Ollama: `ollama run gemma3:27b`
- llama.cpp: `llama-server -hf bartowski/google_gemma-3-27b-it-GGUF:Q4_K_M -c 8192 -ngl 99`
- MLX (Mac): `mlx_lm.chat --model mlx-community/gemma-3-27b-it-4bit`
- Weights: https://huggingface.co/google/gemma-3-27b-it

## Strengths
- Natural, fluent writing and strong multilingual chat
- Image input included
- Small long-context memory cost thanks to sliding-window attention
- Official quantization-aware (QAT) 4-bit weights from Google

## Weaknesses
- No reasoning mode; weak at math and complex coding
- Superseded by Gemma 4 31B and Gemma 4 26B A4B
- Gemma license is more restrictive than Apache 2.0
Full page: https://llamabuilds.ai/models/gemma-3-27b · Hardware guide: https://llamabuilds.ai/models/gemma-3-27b/requirements

---
Source: https://llamabuilds.ai/models/gemma-3-27b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.