Updated July 28, 2026: this guide now includes Gemma 4 12B and the official QAT/Q4 memory guidance. Earlier copy treated the launch family as four models and used one context figure for every checkpoint; both have been corrected.
What is Gemma 4?
Gemma 4 is Google's current open-weight multimodal family for text-and-image workloads. It covers edge-size Mixture-of-Experts models, a practical 12B desktop checkpoint, a sparse 26B-A4B option and a dense 31B flagship.
The useful local-AI question is not “which Gemma is best?” but “which Gemma leaves enough memory for context, image input and the runtime on my machine?”
The current Gemma 4 lineup
| Model | Shape | Official approximate Q4 memory* | Practical tier |
|---|---|---|---|
| Gemma 4 E2B | 2B active MoE | 2.9GB | 8GB systems and edge experiments |
| Gemma 4 E4B | 4B active MoE | 4.5GB | 8-16GB general multimodal use |
| Gemma 4 12B | 12B dense | 6.7GB | 16GB local sweet spot |
| Gemma 4 26B-A4B | 26B total, 4B active MoE | 14.4GB | 24-32GB systems |
| Gemma 4 31B | 31B dense | 17.5GB | 32GB+ workstation |
*Google's published estimates include roughly 20% runtime overhead but not every possible KV-cache, image or application allocation. More context requires more memory.
Which Gemma 4 should you download?
- 8GB: start with E2B; try E4B only with conservative context and enough free memory.
- 16GB: Gemma 4 12B Q4 is the most interesting balance of capability and footprint.
- 24-32GB: compare 26B-A4B against 12B at higher precision or longer context.
- 32GB+: 31B becomes practical, but context and image workloads still need headroom.
QAT, GGUF and local runtimes
Google publishes quantization-aware trained Q4 artifacts for the family. QAT can preserve more quality than converting a floating-point checkpoint after training, but format still matters: a Google QAT release, a GGUF build and an MLX conversion are different artifacts.
LM Studio, llama.cpp-derived runtimes and MLX can all be relevant depending on the checkpoint. Match the model format to the runtime rather than downloading the largest file and assuming every app can load it.
Multi-Token Prediction support
Gemma 4 also has matching MTP drafters for speculative decoding. LM Studio moved MTP support to stable in version 0.4.14. A drafter can improve generation speed without changing the target model's output distribution, but the real gain depends on acceptance rate, context and hardware.
Read the full Gemma 4 MTP guide before enabling it on a production workflow.
Multimodal use
Gemma 4 is designed for unified text and image input. That makes it useful for screenshots, document pages, UI inspection and private visual analysis. Image resolution, tiling and maximum context vary by model and runtime, so use the exact checkpoint card for production limits.
LocalClaw verdict
Gemma 4 12B is the family sweet spot for most 16GB and 32GB users. E2B/E4B win on speed and footprint, 26B-A4B offers an interesting sparse step up, and 31B is the workstation choice. The family is unusually coherent, but no single memory or speed number applies to every checkpoint.