Best GPU for Gemma 3 12B: VRAM Requirements
Gemma 3 12B needs about 30.4 GB of GPU memory in BF16 and 10.3 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.
Gemma 3 12B needs about 30.4 GB of VRAM in FP16, 17 GB in FP8 and 10.3 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on one NVIDIA A100 PCIe 40GB; a 4-bit build fits on one NVIDIA GeForce RTX 5080 16GB.
Gemma 3 12B is a 12.2B-parameter language model from Google DeepMind, released in March 2025, with a 131,072-token context window and published under the Gemma Terms of Use. Its weights alone take 24.4 GB in BF16 and 6.1 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.
Recommended GPUs for Gemma 3 12B
Best for inference
1 GPU × NVIDIA L4 24GB
FP8 or INT8 weights, 17 GB at 8,192 tokens.
Best for fine-tuning
1 GPU × NVIDIA A100 PCIe 40GB
LoRA on BF16 weights, 31.7 GB.
Budget pick
1 GPU × NVIDIA GeForce RTX 5080 16GB
Smallest card class that holds a 4-bit build, 10.3 GB.
Gemma 3 12B VRAM requirements by precision
One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.
| Precision | Weights | KV cache | Total for inference |
|---|---|---|---|
| BF16 / FP16 | 24.4 GB | 3.2 GB | 30.4 GB |
| FP8 | 12.2 GB | 3.2 GB | 17 GB |
| INT4 | 6.1 GB | 3.2 GB | 10.3 GB |
Gemma 3 12B VRAM by task
Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.
| Task | GPU memory |
|---|---|
| Inference at the full 131,072-token context (FP8) | 70.1 GB |
| QLoRA (4-bit base) | 7.9 GB |
| LoRA (BF16 base) | 31.7 GB |
| Full fine-tuning (BF16, AdamW) | 214.7 GB |
GPUs that fit Gemma 3 12B
Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.
| GPU | Memory | BF16 inference (30.4 GB) | FP8 inference (17 GB) | INT4 inference (10.3 GB) | QLoRA (7.9 GB) | LoRA (31.7 GB) |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5080 16GB | 16 GB | 2 GPUs | 2 GPUs | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA GeForce RTX 4090 24GB | 24 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA L4 24GB | 24 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA GeForce RTX 5090 32GB | 32 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA Tesla V100 32GB | 32 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA A100 PCIe 40GB | 40 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA L40S 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA Quadro RTX 8000 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA RTX 6000 Ada 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA RTX A6000 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA A100 PCIe 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA A100 SXM 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H100 PCIe 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H100 SXM 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H200 141GB | 141 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B200 180GB | 180 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B300 288GB | 288 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA GB300 NVL | 288 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
What is Gemma 3 12B?
Gemma 3 12B is the mid-size model of Google's Gemma 3 family, with the same text and image input, support for more than 140 languages and 128K-token context as the 27B model, at less than half its memory. Google also publishes quantization-aware trained INT4 checkpoints.
- Developer
- Google DeepMind
- Released
- March 2025
- Parameters
- 12.2B
- Context window
- 131,072 tokens
- License
- Gemma Terms of Use
Five of every six layers use a 1,024-token sliding window, so the real KV cache at long context is well below this estimate, which counts every layer at full length. The total includes the vision encoder.
Source: Gemma 3 12B model card (as of November 17, 2026).
Frequently asked questions
- How much VRAM does Gemma 3 12B need?
- For inference with an 8,192-token context and one request, Gemma 3 12B needs about 30.4 GB in BF16, 17 GB in FP8, 10.3 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
- Can Gemma 3 12B run on a single GPU?
- Yes. In BF16 it needs about 30.4 GB, which fits one NVIDIA A100 PCIe 40GB (40 GB).
- How much GPU memory does fine-tuning Gemma 3 12B take?
- QLoRA (4-bit base): about 7.9 GB; LoRA (BF16 base): about 31.7 GB; Full fine-tuning (BF16, AdamW): about 214.7 GB. Longer training sequences and larger batches add activation memory on top.
- What is the best GPU for Gemma 3 12B?
- Best for inference: 1 GPU × NVIDIA L4 24GB (FP8 or INT8 weights, 17 GB at 8,192 tokens). Best for fine-tuning: 1 GPU × NVIDIA A100 PCIe 40GB (LoRA on BF16 weights, 31.7 GB). Budget pick: 1 GPU × NVIDIA GeForce RTX 5080 16GB (Smallest card class that holds a 4-bit build, 10.3 GB). Check live offers before booking, as availability changes.
- Can Gemma 3 12B run on an RTX 4090?
- Yes. In FP8 it needs about 17 GB, which fits one RTX 4090 (24 GB). FP16 needs about 30.4 GB, 2 of these cards.
- Can Gemma 3 12B run on an RTX 5090?
- Yes. In FP8 it needs about 17 GB, which fits one RTX 5090 (32 GB). FP16 needs about 30.4 GB, which leaves less than 10 % of the card free.
Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.
Explore the live market
Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.