Google DeepMind · Apache 2.0
Gemma 4
Google’s 2026 family, from phone size to 31B. Reads images and works in 140+ languages.
How much memory it needs
The model file, plus memory for the conversation, plus about 1 GB for the app. Longer conversations need more.
| Size | Version | File | At 8K | At 32K | At 128K |
|---|---|---|---|---|---|
| 5.1B | Q4_K_M | 3.1 GB | 5.2 GB | 5.3 GB | 5.9 GB |
| Q8_0 | 5.0 GB | 7.1 GB | 7.3 GB | 7.9 GB | |
| 8B | Q4_K_M | 5.0 GB | 7.1 GB | 7.5 GB | 9.2 GB |
| Q8_0 | 8.2 GB | 10.3 GB | 10.7 GB | 12.4 GB | |
| 12B | Q4_K_M | 7.1 GB | 8.5 GB | 9.2 GB | 11.8 GB |
| Q8_0 | 12.7 GB | 14.1 GB | 14.7 GB | 17.3 GB | |
| 26B (3.8B active) | Q4_K_M | 17.0 GB | 19.4 GB | 20.1 GB | 22.8 GB |
| Q8_0 | 26.9 GB | 29.3 GB | 29.9 GB | 32.6 GB | |
| 31B | IQ4_XS | 16.4 GB | 19.4 GB | 22.1 GB | 32.7 GB |
| Q4_K_M | 18.3 GB | 21.4 GB | 24.0 GB | 34.6 GB | |
| Q8_0 | 32.6 GB | 35.7 GB | 38.4 GB | 48.9 GB |
Which computers run it
The best size for each machine at an 8K-token conversation. Speeds are estimates from memory bandwidth.
| Computer | Best version | Fit | Speed |
|---|---|---|---|
| Laptop or PC, no graphics card, 32 GB | 26B Q4_K_M | Runs well | 12–25tok/sSmooth |
| RTX 4060 · 8 GB | 8B Q4_K_M | Runs well | 30–50tok/sFast |
| RTX 3060 · 12 GB | 12B Q4_K_M | Runs great | 30–45tok/sSmooth |
| RTX 5060 Ti · 16 GB | 12B Q8_0 | Runs well | 20–35tok/sSmooth |
| RTX 4090 · 24 GB | 31B Q4_K_M | Runs well | 30–50tok/sFast |
| RTX 5090 · 32 GB | 31B Q4_K_M | Runs great | 55–85tok/sFast |
| Mac with M4 · 16 GB | 12B Q4_K_M | Runs well | 9.5–16tok/sReadable |
| Mac with M4 Pro · 48 GB | 31B Q8_0 | Runs well | 4.5–8tok/sReadable |
| Mac with M5 Max · 128 GB | 31B Q8_0 | Runs great | 11–18tok/sSmooth |
| Ryzen AI Max+ 395 · 128 GB | 31B Q4_K_M | Runs great | 8–13tok/sReadable |
| DGX Spark · 128 GB | 31B Q4_K_M | Runs great | 8.5–14tok/sReadable |
| RTX PRO 6000 · 96 GB | 31B Q8_0 | Runs great | 30–50tok/sFast |
| Mac Studio with M5 Ultra · 256 GB | 31B Q8_0 | Runs great | 16–25tok/sSmooth |
| Mac Studio with M3 Ultra · 512 GB | 31B Q8_0 | Runs great | 11–19tok/sSmooth |
Measured speeds
Real runs, not estimates. All measurements.
| Computer | Version | Prompt | Speed | Source |
|---|---|---|---|---|
| Intel Arc B580 Limited Edition 12 GB (Core Ultra 5 250K Plus, CachyOS 7.0.11, oneAPI 2026.0) | Q4_K_M | 512 | 41.34 tok/s | 2026-06-13 |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K | 128K | 43.4 tok/s | 2026-03 |
| NVIDIA GeForce RTX 3090 24 GB (used) | Q4_K | 256K | 64.4 tok/s | 2026-03 |
| NVIDIA GeForce RTX 3090 24 GB (used) | Q4_K | 32K | 31.4 tok/s | 2026-03 |
What it is good and bad at
Good at
- Many languages
- Images in every size; audio in E2B, E4B and 12B
- Very small conversation memory
Not good at
- The 31B is a dense model, so it is slower per word than the 26B-A4B
Start it
Install Ollama, then run one of these in a terminal. LM Studio users can search the same name in the app.
ollama run gemma4:12bollama run gemma4:26b-a4bollama run gemma4:31b