# Gemma 4 26B A4B IT

> OnPrem World · https://onpremworld.com/en/models/gemma-4-26b-a4b · Last updated: 2026-10-14 · CC BY 4.0

A MoE with 3.8B of its 25.2B parameters active: it fits a 16 GB card at 4-bit and generates quickly for its size.

|  |  |
|---|---|
| Developer | Google DeepMind |
| Parameters | 25.2B (3.8B active, MoE) |
| Native context | 262,144 tokens |
| Licence | Apache-2.0 |
| Commercial use | Yes |
| KV cache | 5 global + 25 sliding-window layers (1024 tokens) |
| Released | 2026-03 |

```
ollama run gemma4:26b
```

## Memory needed

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

| Quantization | 4K | 8K | 32K |
|---|---|---|---|
| Q4_K_M | 15.7 GiB | 15.7 GiB | 15.9 GiB |
| Q5_K_M | 18.3 GiB | 18.3 GiB | 18.5 GiB |
| Q6_K | 21.0 GiB | 21.1 GiB | 21.3 GiB |
| Q8_0 | 26.9 GiB | 27.0 GiB | 27.2 GiB |
| FP16 | 50.0 GiB | 50.1 GiB | 50.3 GiB |

## Which hardware runs it (8K)

| Hardware | Memory | Best quant | tok/s (est.) |
|---|---|---|---|
| GeForce RTX 5090 | 32 GB | Q8_0 | 311 |
| GeForce RTX 5080 | 16 GB | Q4_K_M | 292 |
| GeForce RTX 5070 Ti | 16 GB | Q4_K_M | 272 |
| GeForce RTX 4090 | 24 GB | Q6_K | 225 |
| GeForce RTX 4080 SUPER | 16 GB | Q4_K_M | 224 |
| Radeon RX 7900 XTX | 24 GB | Q6_K | 215 |
| GeForce RTX 3090 | 24 GB | Q6_K | 209 |
| Radeon RX 7900 XT | 20 GB | Q5_K_M | 207 |
| Radeon RX 9070 XT | 16 GB | Q4_K_M | 194 |
| RTX 6000 Ada Generation | 48 GB | Q8_0 | 166 |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | FP16 | 165 |
| GeForce RTX 5060 Ti 16GB | 16 GB | Q4_K_M | 136 |
| RTX A6000 | 48 GB | Q8_0 | 133 |
| RTX PRO 5000 72GB Blackwell | 72 GB | FP16 | 124 |
| Radeon AI PRO R9700 | 32 GB | Q8_0 | 111 |
| Arc Pro B70 | 32 GB | Q8_0 | 105 |
| M5 Ultra | 512 GB | FP16 | 103 |
| Arc Pro B60 | 24 GB | Q6_K | 102 |
| Radeon RX 9060 XT 16GB | 16 GB | Q4_K_M | 97 |
| GeForce RTX 4060 Ti 16GB | 16 GB | Q4_K_M | 88 |
| M5 Max (18-core CPU / 32-core GPU) | 36 GB | Q8_0 | 74 |
| M3 Ultra | 512 GB | FP16 | 70 |
| M4 Max (14-core CPU / 32-core GPU) | 36 GB | Q8_0 | 66 |
| M5 Max (18-core CPU / 40-core GPU) | 128 GB | FP16 | 53 |
| M5 Pro | 64 GB | Q8_0 | 49 |
| M4 Max (16-core CPU / 40-core GPU) | 128 GB | FP16 | 47 |
| M4 Pro | 64 GB | Q8_0 | 44 |
| M6 (24GB/32GB configurations, 170 GB/s) | 32 GB | Q6_K | 35 |
| M5 | 32 GB | Q6_K | 32 |
| RTX Spark (N1X) laptops & compact desktops | 128 GB | FP16 | 24 |
| DGX Spark | 128 GB | FP16 | 22 |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) systems | 192 GB | FP16 | 22 |
| Ryzen AI Max+ 395 (Strix Halo) systems | 128 GB | FP16 | 20 |
| GeForce RTX 3060 (12 GB) | 12 GB | Doesn't fit | – |
| GeForce RTX 5070 | 12 GB | Doesn't fit | – |
| Arc B580 | 12 GB | Doesn't fit | – |
| M6 (16GB configuration, 153 GB/s) | 16 GB | Doesn't fit | – |

## Frequently asked questions

### How much VRAM does Gemma 4 26B A4B IT need?

With an 8K context, about 15.7 GiB at Q4_K_M and 27.0 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

### Can Gemma 4 26B A4B IT run on an GeForce RTX 4090?

Yes. Gemma 4 26B A4B IT fits on the GeForce RTX 4090 (24 GB) at Q6_K with an 8K context, using about 21.1 GiB, at an estimated 225 tokens/s.

### Can I run Gemma 4 26B A4B IT on a Mac?

Yes. Gemma 4 26B A4B IT fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 50.1 GiB, at an estimated 70 tokens/s.

### Can Gemma 4 26B A4B IT be used commercially?

Gemma 4 26B A4B 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-4-26B-A4B-it
- https://huggingface.co/google/gemma-4-26B-A4B-it/resolve/main/config.json
- https://huggingface.co/api/models/google/gemma-4-26B-A4B-it?expand[]=safetensors
- https://ollama.com/library/gemma4/tags