# Gemma 4 E4B IT

> OnPrem World · https://onpremworld.com/en/models/gemma-4-e4b · Last updated: 2026-10-14 · CC BY 4.0

A small model built for laptops and other devices: 4.5B effective parameters (8B with per-layer embeddings), with text, image and audio input.

|  |  |
|---|---|
| Developer | Google DeepMind |
| Parameters | 8B |
| Native context | 131,072 tokens |
| Licence | Apache-2.0 |
| Commercial use | Yes |
| KV cache | 7 global + 35 sliding-window layers (512 tokens) |
| Released | 2026-03 |

```
ollama run gemma4:e4b
```

## Memory needed

Weights + KV cache + runtime overhead, in GiB. Lower quantization = less memory, slightly lower quality.

| Quantization | 4K | 8K | 32K |
|---|---|---|---|
| Q4_K_M | 5.4 GiB | 5.5 GiB | 6.2 GiB |
| Q5_K_M | 6.2 GiB | 6.3 GiB | 7.0 GiB |
| Q6_K | 7.1 GiB | 7.2 GiB | 7.9 GiB |
| Q8_0 | 9.0 GiB | 9.1 GiB | 9.7 GiB |
| FP16 | 16.3 GiB | 16.4 GiB | 17.1 GiB |

## Which hardware runs it (8K)

| Hardware | Memory | Best quant | tok/s (est.) |
|---|---|---|---|
| GeForce RTX 5080 | 16 GB | Q8_0 | 79 |
| GeForce RTX 5090 | 32 GB | FP16 | 78 |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | FP16 | 78 |
| GeForce RTX 5070 Ti | 16 GB | Q8_0 | 74 |
| GeForce RTX 4080 SUPER | 16 GB | Q8_0 | 61 |
| RTX PRO 5000 72GB Blackwell | 72 GB | FP16 | 59 |
| GeForce RTX 5070 | 12 GB | Q8_0 | 55 |
| Radeon RX 9070 XT | 16 GB | Q8_0 | 53 |
| M5 Ultra | 512 GB | FP16 | 49 |
| GeForce RTX 4090 | 24 GB | FP16 | 44 |
| Radeon RX 7900 XTX | 24 GB | FP16 | 42 |
| RTX 6000 Ada Generation | 48 GB | FP16 | 42 |
| GeForce RTX 3090 | 24 GB | FP16 | 41 |
| Arc B580 | 12 GB | Q8_0 | 38 |
| GeForce RTX 5060 Ti 16GB | 16 GB | Q8_0 | 37 |
| Radeon RX 7900 XT | 20 GB | FP16 | 35 |
| RTX A6000 | 48 GB | FP16 | 34 |
| M3 Ultra | 512 GB | FP16 | 33 |
| GeForce RTX 3060 (12 GB) | 12 GB | Q8_0 | 30 |
| Radeon AI PRO R9700 | 32 GB | FP16 | 28 |
| Arc Pro B70 | 32 GB | FP16 | 27 |
| Radeon RX 9060 XT 16GB | 16 GB | Q8_0 | 26 |
| M5 Max (18-core CPU / 40-core GPU) | 128 GB | FP16 | 25 |
| GeForce RTX 4060 Ti 16GB | 16 GB | Q8_0 | 24 |
| M4 Max (16-core CPU / 40-core GPU) | 128 GB | FP16 | 22 |
| Arc Pro B60 | 24 GB | FP16 | 20 |
| M5 Max (18-core CPU / 32-core GPU) | 36 GB | FP16 | 19 |
| M4 Max (14-core CPU / 32-core GPU) | 36 GB | FP16 | 17 |
| M5 Pro | 64 GB | FP16 | 12 |
| M6 (16GB configuration, 153 GB/s) | 16 GB | Q8_0 | 12 |
| RTX Spark (N1X) laptops & compact desktops | 128 GB | FP16 | 11 |
| M4 Pro | 64 GB | FP16 | 11 |
| DGX Spark | 128 GB | FP16 | 10 |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) systems | 192 GB | FP16 | 10 |
| Ryzen AI Max+ 395 (Strix Halo) systems | 128 GB | FP16 | 9.6 |
| M6 (24GB/32GB configurations, 170 GB/s) | 32 GB | FP16 | 6.9 |
| M5 | 32 GB | FP16 | 6.2 |

## Frequently asked questions

### How much VRAM does Gemma 4 E4B IT need?

With an 8K context, about 5.5 GiB at Q4_K_M and 9.1 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

### Can Gemma 4 E4B IT run on an GeForce RTX 4090?

Yes. Gemma 4 E4B IT fits on the GeForce RTX 4090 (24 GB) at FP16 with an 8K context, using about 16.4 GiB, at an estimated 44 tokens/s.

### Can I run Gemma 4 E4B IT on a Mac?

Yes. Gemma 4 E4B IT fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 16.4 GiB, at an estimated 33 tokens/s.

### Can Gemma 4 E4B IT be used commercially?

Gemma 4 E4B IT is released under Apache-2.0. The licence allows commercial use (read its conditions).

## Sources

- https://ai.google.dev/gemma/docs/releases
- https://ai.google.dev/gemma/docs/gemma_4_license
- https://huggingface.co/google/gemma-4-E4B-it
- https://huggingface.co/google/gemma-4-E4B-it/resolve/main/config.json
- https://huggingface.co/api/models/google/gemma-4-E4B-it?expand[]=safetensors
- https://ollama.com/library/gemma4/tags