Google DeepMind
Gemma 3 — all variants, VRAM & hardware
Google's Gemma 3 family — dense 1B/4B/12B/27B with multimodal capabilities and Apache-2-compatible license. Excellent efficiency on consumer hardware.
Variant comparison
| Variant | Params | Active | VRAM @ Q4 | Architecture | |
|---|---|---|---|---|---|
| Gemma 3 1B | 1B | 1B | ~0.6 GB | dense | Details → |
| Gemma 3 4B | 4B | 4B | ~2.4 GB | dense | Details → |
| Gemma 3 12B | 12B | 12B | ~7.3 GB | dense | Details → |
| Gemma 3 27B | 27B | 27B | ~16.5 GB | dense | Details → |
Deep-dive guides
Gemma 3 VRAM & hardware guides
FAQ
Frequently asked questions
What is Gemma 3?
Gemma 3 is Google DeepMind's google's gemma 3 family — dense 1b/4b/12b/27b with multimodal capabilities and apache-2-compatible license. excellent efficiency on consumer hardware.
How many variants are in the Gemma 3 family?
We track 4 canonical Gemma 3 variants in our catalog, spanning 1B to 27B parameters.
Which Gemma 3 variant should I run?
Pick the largest variant that fits on your hardware at Q4_K_M quantization. For 8 GB GPUs, the smallest dense variants. For 24 GB GPUs, the 27B dense or 35B MoE (if available). For workstation / Mac Studio, the 122B or larger MoE. Use our fit calculator on each model page for the exact verdict.
See also