Skip to content
local-ai

Gemma 4 26B-A4B

Google · Text generation · MoE 26B (A3.8B) · 262k context · Released 2 April 2026

Commercial use permitted Text + Image Text Open weights Runs on CPU Apple Silicon

Gemma 4 26B-A4B is the mixture-of-experts member of Google's Gemma 4 family. Its sparse design activates roughly 3.8 billion of its 25-or-so billion parameters per token, which keeps it quick and lets it fit a single mid-range card at a 4-bit quantisation, while a 256k context suits long documents. It is multimodal on input, reading images alongside text, and produces text. The family also includes the smaller on-device E2B and E4B, a 12B, and a 31B dense model; this MoE is the sweet spot for capable local use. The licence is the headline for many: Gemma 4 is Apache 2.0, dropping the custom Gemma Terms of Use that earlier generations carried.

Strengths

  • Sparse mixture-of-experts: only about 3.8B of 26B parameters compute per token
  • Fits a single 16GB card at a 4-bit quantisation, with a very long 256k context
  • Apache 2.0, a genuinely permissive licence with no Gemma-specific conditions
  • Reads images as well as text

Weaknesses

  • A mixture-of-experts still needs memory for all its parameters, not only the active ones
  • Independent benchmark results are still accumulating
  • Image support in local runtimes is less mature than text-only use
  • For the strongest single-GPU dense option, the larger Gemma 4 31B may edge it on some tasks

Hardware requirements

Quantisation Approx. VRAM Notes
Q4_K_M ~15GB Common default; fits a 16GB card. Approximate, and rises with context
Q8_0 ~27GB Near-lossless, needs a 32GB card or more. Approximate
BF16 ~52GB Full precision, server or multi-GPU territory. Approximate

Also runs on CPU (slower). Optimised builds available for Apple Silicon.

What you'd need to run this

Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.

Minimum to run it

Q4_K_M · ~15GB needed

One 24GB GPU

NVIDIA Tesla P40

or a Mac or mini-PC with unified memory, if you prefer no discrete GPU, Mac mini M5 Pro .

32–64GB of system RAM alongside the card.

around £700–£1,100

What else 24GB runs →

For good quality

Q8_0 · ~27GB needed

64GB of unified memory

Mac mini M4 Pro

or a single high-memory workstation card, NVIDIA RTX 6000 Ada Generation .

Unified memory is shared with the model, so it is already counted above.

around £1,100–£2,000

See the Affordable Mac build build What else 64GB runs →

Best quality

BF16 · ~52GB needed

64GB of unified memory

Mac mini M4 Pro

or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA A100 80GB .

Unified memory is shared with the model, so it is already counted above.

around £1,100–£2,000

See the Affordable Mac build build What else 64GB runs →

Licence

Apache 2.0 — read the licence

Gemma 4 26B-A4B: common questions

What hardware do I need to run Gemma 4 26B-A4B?
At its most compressed (Q4_K_M) it needs roughly 15GB of VRAM, and about 20GB for good quality. VRAM figures are approximate and depend on context length and settings.
Is Gemma 4 26B-A4B free for commercial use?
Yes. Gemma 4 26B-A4B is licensed under Apache 2.0, which permits commercial use with no meaningful conditions.
Can I run Gemma 4 26B-A4B on Apple Silicon?
Yes. Gemma 4 26B-A4B has builds optimised for Apple Silicon, through MLX or GGUF on a Mac.
Does Gemma 4 26B-A4B run on CPU?
Yes, Gemma 4 26B-A4B can run on the CPU, though generation is slower than on a GPU.
What is Gemma 4 26B-A4B's context window?
Gemma 4 26B-A4B has a context window of 262,144 tokens, about 262k.

Availability

Recommended for

  • A capable, permissively licensed general model on a 16GB to 24GB card
  • Long-document work that needs a large context
  • Local image-and-text understanding under Apache 2.0

Related models

Run it with

Related guides

Glossary

Our coverage

Catalogue entry last verified 2 October 2026. Specifications change; verify anything you are about to spend money on.