Hardware guide · October 2026
What GPU do you need to run Gemma 3 12B?
Gemma 3 12B is a 12.2B dense model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run Gemma 3 12B at Q4_K_M with 8K tokens of context you need about 9GB of VRAM (7.3GB of weights plus KV cache and overhead). The cheapest way in is a 8GB card like the RTX 5060 (~$450). For the best speed per dollar, get the RTX 3060 12GB (~29 tokens/sec). On a Mac or other unified-memory system, the Ryzen AI Max+ 395 runs it at ~15 tokens/sec.
- VRAM at Q4_K_M
- ~9GB
- GPUs that run it
- 36
- Parameters
- 12.2B
- CPU only (est.)
- ~6.0 tok/s
Our recommendations
The best GPUs for Gemma 3 12B
Minimum
RTX 50608GB · ~$450
- Quant
- Q3_K_M
- Speed (est.)
- ~58 tok/s
The smallest card that runs it at Q3_K_M with 8K context.
Best value
RTX 3060 12GB12GB · ~$270 used
- Quant
- Q5_K_M
- Speed (est.)
- ~29 tok/s
The most tokens per second per dollar among cards with room to spare.
Fastest
RTX 509032GB · ~$4,300
- Quant
- Q8_0
- Speed (est.)
- ~105 tok/s
The fastest single GPU in our database for this model.
Mac / unified memory
Ryzen AI Max+ 395128GB unified · ~$3,300
- Quant
- Q8_0
- Speed (est.)
- ~15 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
Gemma 3 12B VRAM requirements
Memory needed at each quantization level. Q4_K_M is the best balance of size and quality for most people; move up to Q5 or Q6 if you have the room.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| Q2_K | Rough | 4.5 GB | 6.1 GB | 7.7 GB |
| Q3_K_M | Usable | 5.9 GB | 7.5 GB | 9.1 GB |
| Q4_K_M | RecommendedDefault | 7.3 GB | 9.0 GB | 11 GB |
| Q5_K_M | Very good | 8.6 GB | 10 GB | 12 GB |
| Q6_K | Near-lossless | 9.9 GB | 12 GB | 13 GB |
| Q8_0 | Practically lossless | 13 GB | 14 GB | 16 GB |
| BF16 | Full precision | 24 GB | 26 GB | 27 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At Q4_K_M, with a 16-bit KV cache, Gemma 3 12B needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.6 GB | 8.7 GB | 8.4 GB |
| 8K tokens | 0.8 GB | 9.0 GB | 8.5 GB |
| 16K tokens | 1.4 GB | 9.5 GB | 8.8 GB |
| 32K tokens | 2.4 GB | 10.6 GB | 9.3 GB |
| 64K tokens | 4.6 GB | 12.7 GB | 10.4 GB |
| 128K tokens | 8.9 GB | 17.0 GB | 12.6 GB |
48 layers in a 5:1 pattern of 1,024-token sliding-window and global layers. Only the 8 global layers grow with context: 2 x 8 x 8 KV heads x 256 dims x 2 bytes = 65,536 bytes per token, about 8.6GB at the full 128K. The 40 sliding layers add a fixed buffer of about 0.3GB.
Compatibility
Every GPU that runs Gemma 3 12B
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| RTX PRO 6000 | 96GB | Q8_0 | ~105 tok/s | Runs well | ~$16,000 |
| RTX 5090 | 32GB | Q8_0 | ~105 tok/s | Runs well | ~$4,300 |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~79 tok/s | Runs well | ~$9,100 |
| RTX 5080 | 16GB | Q6_K | ~72 tok/s | Runs well | ~$1,600 |
| RTX 5070 Ti | 16GB | Q6_K | ~67 tok/s | Runs well | ~$1,200 |
| RTX 4090 | 24GB | Q8_0 | ~61 tok/s | Runs well | ~$2,500 used |
| RTX 5070 | 12GB | Q5_K_M | ~59 tok/s | Runs well | ~$800 |
| RTX 6000 Ada | 48GB | Q8_0 | ~58 tok/s | Runs well | ~$4,750 used |
| RTX 4080 Super | 16GB | Q6_K | ~56 tok/s | Runs well | ~$1,150 used |
| RTX 4070 Ti Super | 16GB | Q6_K | ~51 tok/s | Runs well | ~$900 used |
| RTX 3090 | 24GB | Q8_0 | ~51 tok/s | Runs well | ~$1,050 used |
| M5 Ultra 512GB | 512GB unified | Q8_0 | ~50 tok/s | Runs well | — |
| M5 Ultra 256GB | 256GB unified | Q8_0 | ~50 tok/s | Runs well | ~$9,499 |
| RX 7900 XTX | 24GB | Q8_0 | ~46 tok/s | Runs well | ~$1,050 used |
| RTX 4070 | 12GB | Q5_K_M | ~45 tok/s | Runs well | ~$560 used |
| V100 32GB | 32GB | Q8_0 | ~44 tok/s | Runs well | ~$700 used |
| RTX A6000 | 48GB | Q8_0 | ~42 tok/s | Runs well | ~$4,400 used |
| RX 9070 XT | 16GB | Q6_K | ~39 tok/s | Runs well | ~$800 |
| RTX 5060 Ti 16GB | 16GB | Q6_K | ~35 tok/s | Runs well | ~$620 |
| MI50 32GB | 32GB | Q8_0 | ~34 tok/s | Runs well | ~$600 used |
| M3 Ultra 512GB | 512GB unified | Q8_0 | ~34 tok/s | Runs well | $9,499 |
| R9700 | 32GB | Q8_0 | ~32 tok/s | Runs well | ~$1,800 |
| M5 Max 128GB | 128GB unified | Q8_0 | ~30 tok/s | Runs well | ~$5,100 |
| RTX 3060 12GB | 12GB | Q5_K_M | ~29 tok/s | Runs well | ~$270 used |
| Arc B580 | 12GB | Q5_K_M | ~28 tok/s | Runs well | ~$330 |
| M4 Max 128GB | 128GB unified | Q8_0 | ~27 tok/s | Runs well | $3,499 |
| Arc Pro B70 | 32GB | Q8_0 | ~25 tok/s | Runs well | ~$1,300 |
| RTX 4060 Ti 16GB | 16GB | Q6_K | ~23 tok/s | Runs well | ~$380 used |
| RTX 4000 Ada | 20GB | Q8_0 | ~23 tok/s | Runs well | $1,250 |
| RX 9060 XT 16GB | 16GB | Q6_K | ~20 tok/s | Runs well | ~$470 |
| Arc Pro B60 | 24GB | Q8_0 | ~19 tok/s | Runs well | ~$650 |
| Ryzen AI Max+ PRO 495 | 192GB unified | Q8_0 | ~16 tok/s | Runs well | ~$6,800 |
| DGX Spark | 128GB unified | Q8_0 | ~16 tok/s | Runs well | ~$6,950 |
| Ryzen AI Max+ 395 | 128GB unified | Q8_0 | ~15 tok/s | Runs well | ~$3,300 |
| Tesla P40 | 24GB | Q8_0 | ~15 tok/s | Runs well | ~$230 used |
| RTX 5060 | 8GB | Q3_K_M | ~58 tok/s | Fits, tight | ~$450 |
Tips
Getting the most out of Gemma 3 12B
- Start at Q4_K_M. It keeps nearly all of the quality at about a third of the 16-bit size. Step up to Q5_K_M or Q6_K only if it still fits with the context you need.
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Avoid heavy CPU offload. For a dense model, every token reads every weight. Once a large share of the model sits in system RAM, speed falls off a cliff. A second GPU usually beats offloading.
- Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the Gemma 3 12B overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run Gemma 3 12B?
About 9GB at Q4_K_M with an 8K context window. At Q8_0 (near-lossless) it needs about 14GB, and at Q3_K_M about 7.5GB with some quality loss. Longer contexts add 0.10GB per 1K tokens with a 16-bit KV cache.
Can I run Gemma 3 12B on a 8GB GPU?
Yes. A 8GB card like the RTX 5060 runs it at Q3_K_M at ~58 tokens/sec.
Can I run Gemma 3 12B on a Mac?
Yes. With 128GB of unified memory or more (for example the Ryzen AI Max+ 395), macOS can give the GPU enough memory to run it at Q8_0, at about ~15 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run Gemma 3 12B on CPU only, without a GPU?
Yes, if you have at least 32GB of system RAM. On a desktop with dual-channel DDR5, expect roughly 6.0 tokens/sec. A GPU is strongly recommended for interactive use.