# Gemma 4 26B A4B: hardware requirements
> Gemma 4 26B A4B (Google) is a 26B (3.8B active) mixture-of-experts model with a 256K-token context window, released under the Apache 2.0 license. It needs about 17GB of memory at Q4_K_M with 8K context.
Gemma 4 26B A4B is Google's mixture-of-experts Gemma, with 26B total parameters and 3.8B active (128 experts, 8 used per token), image input, a 256K context, and a thinking mode. Because so few parameters are active, it generates text at small-model speed while matching much larger dense models on reasoning benchmarks. At Q4_K_M it needs about 17GB with 8K context, so it fits a 24GB GPU with plenty of room for context; on 16GB cards it runs at Q3_K_M or with a few experts in system RAM. It replaces Gemma 3 27B for most local uses.

## Recommended hardware
- **Minimum GPU:** RTX 4060 Ti 16GB (16GB), Q3_K_M, ~63 tokens/sec
- **Best value:** RTX 3090 (24GB), Q5_K_M, ~110 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q6_K, ~165 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~26 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 9.8GB | 11.1GB | 11.6GB |
| Q3_K_M | 12.7GB | 14.1GB | 14.6GB |
| Q4_K_M | 15.8GB | 17.3GB | 17.8GB |
| Q5_K_M | 18.5GB | 20.1GB | 20.7GB |
| Q6_K | 21.4GB | 23.2GB | 23.7GB |
| Q8_0 | 27.6GB | 29.5GB | 30.0GB |
| BF16 | 52.0GB | 54.6GB | 55.1GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q6_K | ~165 tok/s | ~$4,300 |
| RTX 5080 | 16GB | Q3_K_M | ~155 tok/s | ~$1,600 |
| RTX 5070 Ti | 16GB | Q3_K_M | ~145 tok/s | ~$1,200 |
| RTX PRO 6000 | 96GB | Q8_0 | ~140 tok/s | ~$16,000 |
| RTX 4080 Super | 16GB | Q3_K_M | ~130 tok/s | ~$1,150 used |
| RTX 4090 | 24GB | Q5_K_M | ~125 tok/s | ~$2,500 used |
| RTX 4070 Ti Super | 16GB | Q3_K_M | ~120 tok/s | ~$900 used |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~115 tok/s | ~$9,100 |
| RTX 3090 | 24GB | Q5_K_M | ~110 tok/s | ~$1,050 used |
| RX 7900 XTX | 24GB | Q5_K_M | ~91 tok/s | ~$1,050 used |
| RTX 5060 Ti 16GB | 16GB | Q3_K_M | ~90 tok/s | ~$620 |
| RX 9070 XT | 16GB | Q3_K_M | ~90 tok/s | ~$800 |
| RTX 6000 Ada | 48GB | Q8_0 | ~89 tok/s | ~$4,750 used |
| V100 32GB | 32GB | Q6_K | ~84 tok/s | ~$700 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~73 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~73 tok/s | ~$9,499 |
| RTX A6000 | 48GB | Q8_0 | ~68 tok/s | ~$4,400 used |
| RTX 4000 Ada | 20GB | Q4_K_M | ~63 tok/s | $1,250 MSRP |
| RTX 4060 Ti 16GB | 16GB | Q3_K_M | ~63 tok/s | ~$380 used |
| R9700 | 32GB | Q6_K | ~61 tok/s | ~$1,800 |
| MI50 32GB | 32GB | Q6_K | ~61 tok/s | ~$600 used |
| RX 9060 XT 16GB | 16GB | Q3_K_M | ~54 tok/s | ~$470 |
| M3 Ultra 512GB | 512GB | Q8_0 | ~54 tok/s | $9,499 MSRP |
| M5 Max 128GB | 128GB | Q8_0 | ~49 tok/s | ~$5,100 |
| Arc Pro B70 | 32GB | Q6_K | ~48 tok/s | ~$1,300 |

## How to run it
- Ollama: `ollama run gemma4:26b`
- llama.cpp: `llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:Q4_K_M -c 8192 -ngl 99`
- Weights: https://huggingface.co/google/gemma-4-26B-A4B-it

## Strengths
- Small-model speed: 3.8B active parameters per token
- Image input, thinking mode, and 256K context
- Compact KV cache thanks to sliding-window attention
- Apache 2.0 license

## Weaknesses
- Slightly too large for 16GB cards at Q4_K_M once context is added
- Behind Gemma 4 31B and Qwen3.8 27B on hard reasoning and coding
- Less world knowledge than a dense model of the same total size
Full page: https://llamabuilds.ai/models/gemma-4-26b-a4b · Hardware guide: https://llamabuilds.ai/models/gemma-4-26b-a4b/requirements

---
Source: https://llamabuilds.ai/models/gemma-4-26b-a4b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.