Gemma 3 27B is the multimodal pick of the 24GB class: it reads images and documents natively, produces polished multilingual text (Italian included), and — uniquely — Google ships official QAT (quantization-aware trained) builds, so its 4-bit versions lose less quality than post-hoc quants of comparable models. If your workload mixes text with scanned documents, screenshots or photos, this is the local model to test first.
VRAM by quantization level
| Quant | Weights | Fits on |
|---|---|---|
| Q4 (QAT) | ~16–17 GB | 24GB comfortably; official QAT quality |
| Q5/Q6 | ~19–22 GB | 24GB with moderate context |
| Q8_0 | ~29 GB | 32GB or 48GB |
Weights only — add KV-cache (grows with context and concurrency) and ~1–2GB runtime overhead. Formula and cache math in the VRAM guide.
Quick start
Vision works out of the box in Ollama and LM Studio (attach images in the prompt/UI). For document pipelines, image tokens consume context — budget accordingly. On Macs, MLX builds run notably fast.
Expected performance
| Hardware | Generation speed |
|---|---|
| RTX 3090 (Q4 QAT) | ~25–32 tok/s |
| RTX 4090 (Q4 QAT) | ~32–42 tok/s |
| Mac M-series Max (MLX) | ~15–22 tok/s |
Indicative single-user figures (llama.cpp/Ollama class runtimes); multi-user serving via vLLM multiplies total throughput 10–20× through batching.
Tips & gotchas
- Always prefer the official QAT quants — this family's 4-bit quality edge is real and free.
- For scanned-PDF RAG, let Gemma read page images directly instead of fighting OCR errors.
- Permissive for business use with a prohibited-use policy — lighter than Llama's, read it once.
FAQ
Gemma 3 27B or Qwen3.6 27B?
Need vision or document images → Gemma. Need best code and a thinking mode → Qwen. Both are strong in Italian; test on your prompts.
Does vision need special hardware?
No — same GPU, but images consume context tokens, so long multi-image chats want 24GB rather than 16GB.