Gemma 3 12B
Gemma 3 12B is Google’s answer to the mid-size class. It punches above its weight on instruction following and is one of the better options if you want image input from a local model without moving to a much larger package.
Who it is for: 16 GB GPU users who want a step up from 8B quality and an easy path to multimodal use. If your license constraints forbid the Gemma terms, the Apache-licensed Mistral 7B v3 is the practical alternative.
Specifications
| Type | llm |
|---|---|
| Parameters | 12B |
| Context window | 32k |
| License | Gemma Terms of Use |
| VRAM recommended | 16 GB (Q8) GB |
| VRAM minimum | 12 GB (Q4) GB |
Pros
- Strong instruction following for its size
- Multimodal: text + image input
- Good multilingual coverage
Cons
- 32k context is modest for long documents
- License is non-commercial in some interpretations — check the terms