Gemma 3 4B VRAM Requirements
Gemma 3 4B needs about 3.7 GB of VRAM at Q4_K_M with an 8K context: 2.4 GB for weights, 0.3 GB for the KV cache and 1.0 GB of runtime overhead. At Q8_0 it needs 5.6 GB and at full FP16 precision 9.3 GB. The smallest GPU that runs it fully is the RTX 3050 8GB.
3.7 GB VRAM
Gemma 3 4B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 2.4 GB
- KV cache 0.3 GB
- Runtime overhead 1.0 GB
Which GPUs can run it
- RTX 3060 12GBFits~81 tok/s
- RTX 4060 Ti 16GBFits~65 tok/s
- RTX 3090 24GBFits~202 tok/s
- Radeon RX 7900 XTX 24GBFits~176 tok/s
- RTX 4090 24GBFits~216 tok/s
- RTX 5090 32GBFits~363 tok/s
- RTX 6000 Ada 48GBFits~206 tok/s
- A100 80GBFits~406 tok/s
- H100 80GB SXMFits~613 tok/s
- Ryzen AI Max+ 395 128GBFits~54 tok/s
- Apple M4 Max 128GBFits~121 tok/s
- Apple M5 Max 128GBFits~135 tok/s
- RTX PRO 6000 Blackwell 96GBFits~363 tok/s
- NVIDIA DGX Spark 128GBFits~57 tok/s
- Apple M3 Ultra 512GBFits~178 tok/s
Gemma 3 4B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 2.4 GB | 3.7 GB | 4.2 GB | 4 GB |
| Q5_K_M | 5.69 | 2.8 GB | 4.2 GB | 4.6 GB | 6 GB |
| Q6_K | 6.59 | 3.3 GB | 4.6 GB | 5.1 GB | 6 GB |
| Q8_0 | 8.50 | 4.3 GB | 5.6 GB | 6.0 GB | 6 GB |
| FP16 / BF16 | 16.00 | 8.0 GB | 9.3 GB | 9.8 GB | 10 GB |
| Q3_K_M | 3.91 | 2.0 GB | 3.3 GB | 3.7 GB | 4 GB |
| Q2_K | 3.00 | 1.5 GB | 2.8 GB | 3.3 GB | 4 GB |
VRAM by context length
At Q4_K_M. 5 of 34 layers keep the full context. The other 29 use a 1024-token sliding window, so their cache stops growing after 1K tokens.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.2 GB | 0.1 GB | 3.7 GB | 3.6 GB |
| 8K tokens | 0.3 GB | 0.1 GB | 3.7 GB | 3.6 GB |
| 32K tokens | 0.7 GB | 0.4 GB | 4.2 GB | 3.9 GB |
| 128K tokens | 2.6 GB | 1.4 GB | 6.1 GB | 4.9 GB |
GPUs that can run Gemma 3 4B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| RTX 3050 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~51 tok/s |
| RTX 4060 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~62 tok/s |
| RTX 4060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~65 tok/s |
| Radeon RX 7600 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~55 tok/s |
| RTX 3060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~100 tok/s |
| RTX 3070 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~100 tok/s |
| RTX 5060 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~100 tok/s |
| RTX 5060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~100 tok/s |
| RTX 3070 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~134 tok/s |
| Intel Arc B570 10GB | 10.0 GB | Fits | FP16 / BF16 | 128K | ~66 tok/s |
| RTX 3080 10GB | 10.0 GB | Fits | FP16 / BF16 | 128K | ~166 tok/s |
| Apple M4 16GB | 10.7 GB | Fits | FP16 / BF16 | 128K | ~27 tok/s |
| RTX 2080 Ti 11GB | 11.0 GB | Fits | FP16 / BF16 | 128K | ~136 tok/s |
| RTX 2060 12GB | 12.0 GB | Fits | FP16 / BF16 | 128K | ~76 tok/s |
About Gemma 3 4B
A multimodal model with a 1024-token sliding window on five of every six layers. Only five layers keep the full context, so a 128K window costs a fraction of what it would on a standard transformer. Add about 0.8 GB if you load the vision encoder.
| Spec | Value |
|---|---|
| Parameters | 4.3B |
| Layers | 34 |
| Hidden size | 2560 |
| Attention heads / KV heads | 8 / 4 |
| Head dimension | 256 |
| Vocabulary | 262,208 |
| Max context (used here) | 128K tokens |
| Sliding window | 1024 tokens on 29 of 34 layers |
| KV cache per token (FP16) | 20 KB |
| Released | 2025-03 |
| License | Gemma Terms of Use |
Gemma 3 4B VRAM: common questions
Why does Gemma 3 use so little memory for long context?
Most of its layers only attend to the last 1,024 tokens, so only 5 of 34 layers store the full context in the KV cache.
How much VRAM does Gemma 3 4B need?
About 3.7 GB at Q4_K_M with an 8K context, 5.6 GB at Q8_0 and 9.3 GB at FP16. Longer contexts add 0.02 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Gemma 3 4B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the RTX 3050 8GB (8.0 GB usable).