Skip to main content
Local LLM Guide

Model

Gemma 4 26B-A4B Instruct

Start with a supported compressed build and short context. 64GB RAM estimate. Best use: Best for high-RAM desktops and users who want MoE speed.

25.2B total / 3.8B activeMultimodal

Best use

Best for high-RAM desktops and users who want MoE speed.

Memory planning

64GB RAM estimate

Expected pace

Depends on your hardware, app, and conversation length

License

Apache 2.0; read the model’s terms before use

Who should use this model?

  • Best for high-RAM desktops and users who want MoE speed.

Choose something else when

  • Avoid for tiny 8GB machines needing full-quality long context, or projects requiring non-text outputs today.

What hardware makes sense?

The download is only part of the memory needed. Leave room for the app, the input it processes, and other open programs. For chat models, context includes the instructions, conversation, and answer. Longer context needs more memory.

CPU only
Use a smaller model for this memory limit.
8GB system memory
Use a smaller model for this memory limit.
16GB system memory
Use a smaller model for this memory limit.
32GB system memory
Memory or response time may limit use. Try a smaller model first.
8GB GPU memory
Use a smaller model for this memory limit.
12GB GPU memory
Use a smaller model for this memory limit.
24GB GPU memory
Use a smaller model for this memory limit.
Apple unified memory
Memory or response time may limit use. Try a smaller model first.

Model facts that change the decision

Parameters
25.2B total / 3.8B active
Input
Text · Image
Published context limit
256,000 tokens
Model format
safetensors · GGUF · Ollama · MLX
Suggested compression
Q4_K_M for first local attempts unless a cited runtime recommends a different default.
Typical Q4 file
14.4GB before runtime and context overhead
Languages
Unknown

Using image input

This model needs an app or library that supports its image-processing components. Follow the exact model’s instructions for preparing image inputs. Some GGUF downloads also need a matching projector file; loading the text model alone does not establish image support.

First setup

Ollama

ollama run gemma4:26b

LM Studio: Check that your installed app supports the exact model format and task before downloading it.

llama.cpp

Use a cited GGUF build with current llama.cpp-compatible tooling.

Conditions that change the answer

  • Official Q4_0 memory estimates cover model loading only and do not include full runtime overhead, supporting software, or conversation memory.
  • Long 128K and 256K contexts can require much more RAM or VRAM than the base model-size number suggests.
  • Audio input is available only on E2B, E4B, and 12B, not on 26B-A4B or 31B.
  • All variants generate text output; they do not generate image or audio output.
  • The 26B-A4B MoE activates only 3.8B parameters per token but still must load all parameters into memory.
  • Use the official Apache 2.0 license and Google policy links for legal review before treating commercial use as cleared.
  • Commercial use: permitted
  • Gemma 4 is Google DeepMind’s newest open Gemma family for local reasoning, coding, agents, and multimodal chat. It spans tiny edge models, a 12B audio-capable model, a fast MoE, and a stronger 31B dense model.

Read the controlling license text →