# Gemma 4 E4B: hardware requirements
> Gemma 4 E4B (Google) is a 8B dense model with a 128K-token context window, released under the Apache 2.0 license. It needs about 6.3GB of memory at Q4_K_M with 8K context.
Gemma 4 E4B is the larger of Google's two on-device Gemma 4 models: about 4.5B effective parameters (8B including per-layer embeddings) with text, image, and audio input and an optional thinking mode. It replaces Gemma 3n and the small Gemma 3 models, and it ships under Apache 2.0 instead of the old Gemma license. At Q4_K_M it needs about 6GB with 8K context, so it runs on any 8GB GPU, on 16GB laptops and Macs, and even on CPU-only machines at usable speed. Its KV cache is tiny, so the full 128K context adds only about 2GB.

## Recommended hardware
- **Minimum GPU:** RTX 5060 (8GB), Q4_K_M, ~62 tokens/sec
- **Best value:** RTX 5060 (8GB), Q4_K_M, ~62 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q8_0, ~130 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~20 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 3.3GB | 4.2GB | 4.6GB |
| Q3_K_M | 4.3GB | 5.2GB | 5.6GB |
| Q4_K_M | 5.4GB | 6.3GB | 6.7GB |
| Q5_K_M | 6.3GB | 7.3GB | 7.7GB |
| Q6_K | 7.3GB | 8.3GB | 8.7GB |
| Q8_0 | 9.5GB | 10.5GB | 10.9GB |
| BF16 | 17.8GB | 19.1GB | 19.5GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q8_0 | ~130 tok/s | ~$4,300 |
| RTX PRO 6000 | 96GB | Q8_0 | ~130 tok/s | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~100 tok/s | ~$9,100 |
| RTX 4090 | 24GB | Q8_0 | ~79 tok/s | ~$2,500 used |
| RTX 5080 | 16GB | Q8_0 | ~75 tok/s | ~$1,600 |
| RTX 6000 Ada | 48GB | Q8_0 | ~75 tok/s | ~$4,750 used |
| RTX 5070 Ti | 16GB | Q8_0 | ~70 tok/s | ~$1,200 |
| RTX 3090 | 24GB | Q8_0 | ~66 tok/s | ~$1,050 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~64 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~64 tok/s | ~$9,499 |
| RTX 5060 | 8GB | Q4_K_M | ~62 tok/s | ~$450 |
| RX 7900 XTX | 24GB | Q8_0 | ~60 tok/s | ~$1,050 used |
| RTX 4080 Super | 16GB | Q8_0 | ~59 tok/s | ~$1,150 used |
| V100 32GB | 32GB | Q8_0 | ~58 tok/s | ~$700 used |
| RTX A6000 | 48GB | Q8_0 | ~55 tok/s | ~$4,400 used |
| RTX 4070 Ti Super | 16GB | Q8_0 | ~54 tok/s | ~$900 used |
| RTX 5070 | 12GB | Q8_0 | ~54 tok/s | ~$800 |
| M3 Ultra 512GB | 512GB | Q8_0 | ~44 tok/s | $9,499 MSRP |
| MI50 32GB | 32GB | Q8_0 | ~44 tok/s | ~$600 used |
| R9700 | 32GB | Q8_0 | ~41 tok/s | ~$1,800 |
| RX 9070 XT | 16GB | Q8_0 | ~41 tok/s | ~$800 |
| RTX 4070 | 12GB | Q8_0 | ~41 tok/s | ~$560 used |
| M5 Max 128GB | 128GB | Q8_0 | ~40 tok/s | ~$5,100 |
| RTX 5060 Ti 16GB | 16GB | Q8_0 | ~36 tok/s | ~$620 |
| M4 Max 128GB | 128GB | Q8_0 | ~36 tok/s | $3,499 MSRP |

## How to run it
- Ollama: `ollama run gemma4:e4b`
- llama.cpp: `llama-server -hf unsloth/gemma-4-E4B-it-GGUF:Q4_K_M -c 8192 -ngl 99`
- Weights: https://huggingface.co/google/gemma-4-E4B-it

## Strengths
- Text, image, and audio input in a model that fits 8GB
- Very small KV cache: the full 128K context adds about 2GB
- Optional thinking mode for harder questions
- Apache 2.0 license

## Weaknesses
- Limited knowledge and reasoning depth compared with 9B to 12B models
- Per-layer embeddings make the file larger than the 4.5B effective size suggests
- Long thinking traces are slow on CPU-only machines
Full page: https://llamabuilds.ai/models/gemma-4-e4b · Hardware guide: https://llamabuilds.ai/models/gemma-4-e4b/requirements

---
Source: https://llamabuilds.ai/models/gemma-4-e4b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.