Gemma Models: Google's Lightweight Local AI
Google DeepMind's Gemma family delivers impressive quality from compact models. The current Gemma 4 line starts with E4B, a 4.5B on-device model that runs on an 8GB Mac, then steps up to the 26B-A4B MoE (25.2B total, 3.8B active, 16GB load) and the dense 31B (20GB load, 32GB floor), all Apache 2.0, with the 31B taking text and images. If you want a well-tuned, safety-conscious model that fits modest hardware, Gemma is a strong choice.
- developer
- Google DeepMind
- local models
- 19
- size range
- 1B-31B
- RAM range
- 2-48 GB
All Gemma models
19 local models| Model | Size | Quant | VRAM | Min RAM | Best For | Quality | Ollama |
|---|---|---|---|---|---|---|---|
| Gemma 3 1B Instruct | 1B | Q4_K_M | 1 GB | 2 GB | Chat, Mobile | 52 | |
| Gemma 3 1B Instruct (Q8) | 1B | Q8_0 | 1 GB | 4 GB | Chat, Mobile | 54 | |
| Gemma 2 2B Instruct | 2B | Q4_K_M | 1.8 GB | 4 GB | Chat | 58 | |
| Gemma 4 E2B | 2.3B | Q4_K_M | 2.3 GB | 4 GB | IoT, Mobile, Edge | 68 | |
| Gemma 4 E2B (Q8) | 2.3B | Q8_0 | 4.6 GB | 8 GB | IoT, Mobile, Edge | 70 | |
| Gemma 3 4B Instruct | 4B | Q4_K_M | 3.5 GB | 6 GB | Chat, Coding | 72 | |
| Gemma 3 4B Instruct (Q8) | 4B | Q8_0 | 3.9 GB | 8 GB | Chat, Coding | 74 | |
| Gemma 4 E4B | 4.5B | Q4_K_M | 4 GB | 6 GB | On-device, Mobile, Chat | 78 | |
| Gemma 4 E4B (Q8) | 4.5B | Q8_0 | 7.5 GB | 16 GB | On-device, Mobile, Chat | 80 | |
| Gemma 2 9B Instruct | 9B | Q4_K_M | 7 GB | 12 GB | Chat, Coding | 76 | |
| Gemma 3 12B Instruct | 12B | Q4_K_M | 9.5 GB | 16 GB | Chat, Quality | 85 | |
| Gemma 4 12B | 12B | Q4_K_M | 8 GB | 12 GB | Chat, Coding, Multimodal | 90 | |
| Gemma 4 12B (Q8) | 12B | Q8_0 | 12.8 GB | 24 GB | Chat, Coding, Multimodal | 92 | |
| Gemma 4 26B-A4B | 26B | Q4_K_M | 16 GB | 24 GB | Chat, Coding, Multimodal | 92 | |
| Gemma 4 26B-A4B (Q8) | 26B | Q8_0 | 28.1 GB | 48 GB | Chat, Coding, Multimodal | 94 | |
| Gemma 2 27B Instruct | 27B | Q4_K_M | 21 GB | 32 GB | Quality, Coding | 80 | |
| Gemma 3 27B Instruct | 27B | Q4_K_M | 21 GB | 32 GB | Quality, Coding | 85 | |
| Gemma 4 31B | 31B | Q4_K_M | 20 GB | 32 GB | Quality, Coding, Multimodal | 93 | |
| Gemma 4 31B (Q8) | 31B | Q8_0 | 30.9 GB | 48 GB | Quality, Coding, Multimodal | 95 |
Key features
- Gemma 4 E4B: 4.5B effective, runs on 8GB Macs
- Gemma 4 26B-A4B MoE: 25.2B total, 3.8B active
- Gemma 4 31B dense: 20GB load, 32GB RAM floor
- Apache 2.0 across the Gemma 4 line
- Previous Gemma 2 / 3 sizes still on Ollama
Device compatibility
Which Gemma models can run on each device class, based on minimum RAM requirements.
| Model | iPhone | Air | Pro | Studio | Mini | |
|---|---|---|---|---|---|---|
| Gemma 3 1B Instruct (1B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Gemma 3 1B Instruct (Q8) (1B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Gemma 2 2B Instruct (2B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Gemma 4 E2B (2.3B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Gemma 4 E2B (Q8) (2.3B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Gemma 3 4B Instruct (4B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Gemma 3 4B Instruct (Q8) (4B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Gemma 4 E4B (4.5B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Gemma 4 E4B (Q8) (4.5B) | No | Possible | Possible | Excellent | Possible | Possible |
| Gemma 2 9B Instruct (9B) | Possible | Possible | Possible | Excellent | Possible | Possible |
| Gemma 3 12B Instruct (12B) | No | Possible | Possible | Excellent | Possible | Possible |
| Gemma 4 12B (12B) | Possible | Possible | Possible | Excellent | Possible | Possible |
| Gemma 4 12B (Q8) (12B) | No | Possible | Possible | Possible | Possible | Possible |
| Gemma 4 26B-A4B (26B) | No | Possible | Possible | Possible | Possible | Possible |
| Gemma 4 26B-A4B (Q8) (26B) | No | No | Possible | Possible | Possible | No |
| Gemma 2 27B Instruct (27B) | No | Possible | Possible | Possible | Possible | Possible |
| Gemma 3 27B Instruct (27B) | No | Possible | Possible | Possible | Possible | Possible |
| Gemma 4 31B (31B) | No | Possible | Possible | Possible | Possible | Possible |
| Gemma 4 31B (Q8) (31B) | No | No | Possible | Possible | Possible | No |
RAM requirements
Largest dense Q4 model that fits each memory tier
Get told when a better model fits your MacBook Air M5 16 GB
Get one email when a new open-weight model is a better fit for this machine.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Config reports
| Source | Machine | tok/s | Quant | Runtime | Context | Note |
|---|---|---|---|---|---|---|
| Speed test | Apple M6 12-core GPU 32 GBGemma 4 26B-A4B | 59.3 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M1 Max 64 GBGemma 4 26B-A4B | 36.0 | Q8_0 | Ollama | - | 1 run |
Speed test: runs submitted through the ModelFit speed test tool, one median per machine.
Share your config
Frequently asked questions
What is the best Gemma model for a MacBook Air?
Gemma 4 E4B is the current pick for an 8GB Air: `ollama run gemma4:e4b` needs 6GB RAM minimum and loads about 4GB. On a 16GB machine you can move up to the larger Gemma 3 or Gemma 2 sizes, and 24GB+ opens the Gemma 4 26B-A4B MoE.
Can Gemma run on an iPhone?
Yes. Gemma 4 E4B is built for on-device use and runs on high-end iPhones and any 8GB Mac. Quality is basic but useful for simple chat and text tasks.
How does Gemma compare to Llama at similar sizes?
At small sizes they are close. Gemma tends to be more conservative and safety-tuned, while Llama is more flexible and has a larger fine-tune ecosystem. Pick based on whether you want guardrails or freedom.
