Written by Jakub Rusinowski · Last updated March 12, 2025
Yes — comfortably
Yes, comfortably — Gemma 3 12B Instruct at Q8_0 needs about 16.6 GB of the 24 GB usable on RTX 3090 Desktop (24 GB VRAM, 64 GB RAM), leaving ~7.4 GB spare and running at ~44.2 tok/s (estimated), with room for about 16,384 tokens of context.
Confidence: high · Recommended quantization: Q8_0 · Estimated speed: ~44.2 tok/s
| Usable memory for models | 24 GB |
| Memory bandwidth | 936 GB/s |
| Form factor | Desktop |
| Operating system | Windows or Linux |
| Memory upgradeable | Yes |
| Quant | Memory needed | Fits 24 GB? | Max context | Est. speed | Download |
|---|---|---|---|---|---|
| F16 | 27.9 GB | ✗ No | — | — | 24 GB |
| Q8_0 | 16.6 GB | ✓ Yes | 16K | ~44.2 tok/s | 12.8 GB |
| Q6_K | 13.7 GB | ✓ Yes | 32K | ~54.2 tok/s | 9.8 GB |
| Q5_K_M | 12.4 GB | ✓ Yes | 32K | ~60.4 tok/s | 8.5 GB |
| Q4_K_M | 11.1 GB | ✓ Yes | 32K | ~67.7 tok/s | 7.2 GB |
| Q3_K_M | 9 GB | ✓ Yes | 32K | ~85.2 tok/s | 5.1 GB |
| Q2_K | 7.8 GB | ✓ Yes | 32K | ~99.4 tok/s | 3.9 GB |
| Variant | Needs at Q4_K_M | Fits? | Est. speed |
|---|---|---|---|
| Gemma 3 27B Instruct | 25.4 GB | ✗ Too large | — |
| Gemma 3 12B Instruct | 11.1 GB | ✓ Fits | ~67.7 tok/s |
| Gemma 3 4B Instruct | 4.4 GB | ✓ Fits | ~153.9 tok/s |
| Gemma 3 1B Instruct | 2 GB | ✓ Fits | ~282.8 tok/s |
Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving
Yes, comfortably — Gemma 3 12B Instruct at Q8_0 needs about 16.6 GB of the 24 GB usable on RTX 3090 Desktop (24 GB VRAM, 64 GB RAM), leaving ~7.4 GB spare and running at ~44.2 tok/s (estimated), with room for about 16,384 tokens of context.
Q8_0 — it needs about 16.6 GB of the 24 GB available, downloads as roughly 12.8 GB, and runs at an estimated 44.2 tokens/sec with up to 16K of context.
Nothing binding — the model fits with headroom and generates at a usable speed on this hardware.
Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving