Gemma 4 12B
VisionGoogle · 12B · Dense
Gemma 4 mid-size base model (official)
0
0
2026-04
256K tokens
Use cases
Your hardware
16 GB Reference Desktop
Quantization options
What is quantization?| Quant | Bits | VRAM | Disk | Quality | Status |
|---|---|---|---|---|---|
| Q2_K | 2 | 4.3 GB | 3.7 GB | low | Runs great · ~47 t/s |
| Q3_K_M | 3 | 5.9 GB | 5.1 GB | moderate | Runs well · ~34 t/s |
| Q4_K_M | 4 | 6.6 GB | 5.9 GB | good | Runs well · ~31 t/s |
| Q5_K_M | 5 | 8.2 GB | 7.3 GB | good | Decent · ~25 t/s |
| Q6_K | 6 | 9.7 GB | 8.8 GB | excellent | Tight fit · ~21 t/s |
| Q8_0 | 8 | 12.8 GB | 11.7 GB | excellent | Tight fit · ~16 t/s |
| F16 | 16 | 25.1 GB | 23.5 GB | lossless | Barely runs · ~3 t/s |
About this model
Gemma 4 12B (Gemma 4 12b) is a 12B dense transformer from Google in the Gemma family, designed primarily for vision workflows. Supports up to 256K context tokens; start at 4K–8K with Q8_0 KV cache on tight GPUs.
Hardware planning baselines start at 6.7 GB minimum memory and 11.2 GB recommended memory. At Q4_K_M, weights plus baseline runtime overhead use roughly 6.6 GB of RAM/VRAM and 5.9 GB of disk space. As a 12B dense checkpoint, every parameter participates in each token forward pass, meaning decoding speed scales linearly with GPU memory bandwidth and drops sharply if layers spill over PCIe into host DDR5 RAM.
Deploy Gemma 4 12B locally via llama.cpp with Flash Attention enabled. Under Gemma Terms of Use, review official weight cards before commercial redistribution.
Highlights
- Local Fit Baseline6.7 GB minimum and 11.2 GB recommended RAM/VRAM.
- Primary QuantQ4_K_M targets 6.6 GB memory and 5.9 GB disk.
- Context & Capabilities256K max context · Dense.
System requirements by device
Memory planning targets 6.7 GB minimum RAM/VRAM and 11.2 GB recommended memory. The baseline quant is Q4_K_M, estimated at 6.6 GB memory and 5.9 GB disk. Actual runtime usage shifts with context length, KV cache precision, and OS background overhead.
| Device | Notes |
|---|---|
| Mac (Apple silicon) | Use unified memory headroom (~75% physical RAM max) as the primary fit signal. Start with Q4_K_M and close heavy browser tabs first. |
| Windows / Linux GPU | Keep weights plus KV cache inside dedicated VRAM so Windows WDDM does not spill layers over PCIe into Shared GPU Memory. |
| Android | Feasible on 8–16 GB phones via AI Edge Gallery or PocketPal when keeping context short (2K–4K) and closing background apps. |
| iPhone / iPad | Target 8 GB+ Pro iPhones or M-series iPads with Q4_K_M and 2K–4K context so allocations stay below iOS Jetsam ceilings. |
Start with a short session and leave memory for the operating system. A file fitting on disk does not guarantee that it will load in memory.
Estimated memory by quantization
These are rough estimates, not tested minimum requirements. Check the actual file and runtime documentation; context, offloading and operating-system overhead can change the memory needed.
| Quant Tier | Estimated memory | Disk Size | Quality |
|---|---|---|---|
| Q2_K | 4.3 GB | 3.7 GB | low |
| Q3_K_M | 5.9 GB | 5.1 GB | moderate |
| Q4_K_M | 6.6 GB | 5.9 GB | good |
| Q5_K_M | 8.2 GB | 7.3 GB | good |
| Q6_K | 9.7 GB | 8.8 GB | excellent |
| Q8_0 | 12.8 GB | 11.7 GB | excellent |
| F16 | 25.1 GB | 23.5 GB | lossless |
How to run Gemma 4 12B on your device
Start with a short prompt and the smallest quant that still gives the quality you need. A clean first run is more useful than loading a larger file with no memory headroom.
Run Gemma 4 12B on Mac
- Install LM Studio or llama.cpp on macOS Apple Silicon.
- Pull Q4_K_M (6.6 GB) first and confirm it stays below ~75% of your Mac's unified memory.
- Enable Flash Attention and cap initial context at 8K until memory pressure stays below 11.2 GB.
- Consult the Mac guide for Activity Monitor tuning and Metal offload tips.
Run Gemma 4 12B on Windows
- Install LM Studio for Windows for visual VRAM inspection, or Ollama when you want a localhost API endpoint.
- Load Q4_K_M (~6.6 GB) and verify Dedicated GPU Memory stays within VRAM without spilling into Shared GPU Memory.
- Reserve at least 6.7 GB memory minimum and target 11.2 GB for comfortable context growth.
- Enable Q8_0 KV cache or drop one quant tier if generation slows down over PCIe.
Run Gemma 4 12B on Android
- Use AI Edge Gallery or PocketPal AI on an Android phone with at least 8–12 GB RAM.
- Select Q4_K_M or a smaller mobile-sized quant (~6.6 GB) and keep context at 2K–4K.
- Close background camera and browser apps before loading so Android LMK does not kill the process.
- Consult the Android guide for SoC thermal and NPU/GPU delegate tips.
Run Gemma 4 12B on iPhone or iPad
- Install PocketPal or a Metal-enabled llama.cpp iOS client on an 8 GB+ iPhone or M-series iPad.
- Import Q4_K_M via Files and keep initial context at 2048 tokens to stay beneath iOS Jetsam limits.
- Test one short text prompt before running multi-turn chats.
- Move to a Mac if iOS terminates the app due to memory pressure.
Gemma 4 12B vs alternatives
| Model | Params | Context | Min RAM | License |
|---|---|---|---|---|
| Gemma 4 12B | 12B | 256K | 6.7 GB | Gemma |
| Gemma 3 12B | 12B | 128K | 6.7 GB | Gemma |
| Gemma 4 12B IT | 12B | 256K | 6.7 GB | Apache 2.0 |
| Gemma 2 9B | 9B | 8K | 5 GB | Gemma |
| Gemma 4 E4B IT | 8B | 128K | 4.5 GB | Apache 2.0 |
| Gemma 4 E4B | 8B | 128K | 4.5 GB | Gemma |
Use this table to choose the smallest nearby model that still matches the job. A related model with lower RAM can be the better local choice when the current page fits only at the edge of available memory.
Frequently asked questions
- What is Gemma 4 12B?
- Gemma 4 12B is a 12B dense transformer from Google, supporting up to 256K tokens with 6.7 GB minimum RAM and 11.2 GB recommended RAM.
- How do I install Gemma 4 12b on Windows?
- Use LM Studio on Windows for the first fit test, load Q4_K_M, and watch RAM plus GPU memory before raising context or switching to a larger quant.
- How much VRAM or RAM does Gemma 4 12B require?
- At Q4_K_M, Gemma 4 12B occupies roughly 6.6 GB of active memory and 5.9 GB on disk. Add 2–6 GB of spare headroom for KV cache buffers, runtime scratch space, and OS apps (11.2 GB recommended).
- What is the best way to run Gemma 4 12B locally?
- Run Gemma 4 12B in llama.cpp starting with Q4_K_M. Keep initial context at 4K–8K tokens and confirm 100% GPU/Metal offload before scaling context length.
- Can Gemma 4 12B run on Android or iPhone?
- Yes, on flagship 8–16 GB Android phones (via AI Edge Gallery or PocketPal) or 8 GB+ iPhones when capped to Q4_K_M and short 2K–4K context windows.
- How does Gemma 4 12B compare to Gemma 3 12B?
- Gemma 3 12B (12B, 6.7 GB min RAM) sits closest in size. Compare active parameters, context ceiling (256K vs 128K), and license terms (Gemma) before switching.