About this model
Gemma 3 12B is Google's multimodal-capable mid-size model: it reads images as well as text, and its quality sits comfortably between 8B and 14B peers. It is a common choice for 12 GB cards that want a vision-capable model.
What it is good for
- Image understanding plus chat in one local model
- Document and screenshot analysis
- Multilingual writing tasks
Limitations
- Vision inputs consume additional memory beyond text
- Not a specialist coder; use a coder-focused model for programming
Memory by quantization (estimates)
Weights at a given bit-width plus estimated runtime and context overhead. These are estimates — see the methodology.
| Quantization | Approx. memory | Fits a 8 GB card | Recommended class |
|---|---|---|---|
| Q4 (GGUF) | ~7.3 GB | Good | 8 GB+ |
| Q5 (GGUF) | ~8.5 GB | Tight | 9 GB+ |
| Q8 (GGUF) | ~12.8 GB | Tight | 13 GB+ |
GPUs that can run it
FAQ
Can Gemma 3 12B run on 8 GB VRAM?
Yes at Q4 — roughly 7–8 GB including text-only usage. Vision inputs add memory, so keep images modest on 8 GB.
Does it really process images locally?
Yes — Gemma 3 is multimodal; runtimes like Ollama and llama.cpp support its vision input at appropriate quantization.
How does it compare to Qwen2.5 14B?
Similar class. Gemma adds vision; Qwen 14B is often stronger at tool use and structured output. Both need ~12 GB at comfortable quantization.
Can it run on CPU?
At Q4, yes — with about 8–10 GB of RAM. Expect CPU-class speeds (a few tok/s to low tens depending on the machine).