Gemma 4 E2B VRAM Requirements
Gemma 4 E2B needs about 4.0 GB of VRAM at Q4_K_M with an 8K context: 2.9 GB for weights, 54 MB for the KV cache and 1.0 GB of runtime overhead. At Q8_0 it needs 6.1 GB and at full FP16 precision 10.6 GB. The smallest GPU that runs it fully is the RTX 3050 8GB.
4.0 GB VRAM
Gemma 4 E2B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 2.9 GB
- KV cache 54 MB
- Runtime overhead 1.0 GB
Which GPUs can run it
- RTX 3060 12GBFits~73 tok/s
- RTX 4060 Ti 16GBFits~59 tok/s
- RTX 3090 24GBFits~182 tok/s
- Radeon RX 7900 XTX 24GBFits~160 tok/s
- RTX 4090 24GBFits~195 tok/s
- RTX 5090 32GBFits~330 tok/s
- RTX 6000 Ada 48GBFits~187 tok/s
- A100 80GBFits~369 tok/s
- H100 80GB SXMFits~560 tok/s
- Ryzen AI Max+ 395 128GBFits~48 tok/s
- Apple M4 Max 128GBFits~109 tok/s
- Apple M5 Max 128GBFits~122 tok/s
- RTX PRO 6000 Blackwell 96GBFits~330 tok/s
- NVIDIA DGX Spark 128GBFits~51 tok/s
- Apple M3 Ultra 512GBFits~161 tok/s
Gemma 4 E2B 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 | 2.9 GB | 4.0 GB | 4.1 GB | 6 GB |
| Q5_K_M | 5.69 | 3.4 GB | 4.5 GB | 4.6 GB | 6 GB |
| Q6_K | 6.59 | 3.9 GB | 5.0 GB | 5.1 GB | 6 GB |
| Q8_0 | 8.50 | 5.0 GB | 6.1 GB | 6.3 GB | 8 GB |
| FP16 / BF16 | 16.00 | 9.5 GB | 10.6 GB | 10.7 GB | 12 GB |
| Q3_K_M | 3.91 | 2.3 GB | 3.4 GB | 3.5 GB | 4 GB |
| Q2_K | 3.00 | 1.8 GB | 2.9 GB | 3.0 GB | 4 GB |
VRAM by context length
At Q4_K_M. Only 15 of 35 layers have their own KV cache; the last 20 reuse caches from earlier layers. Of those 15, 12 use a 512-token sliding window and 3 keep the full context (1 KV head, 512-dimension keys and values), so the cache is tiny.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 30 MB | 16 MB | 3.9 GB | 3.9 GB |
| 8K tokens | 54 MB | 29 MB | 4.0 GB | 3.9 GB |
| 32K tokens | 0.2 GB | 0.1 GB | 4.1 GB | 4.0 GB |
| 128K tokens | 0.8 GB | 0.4 GB | 4.7 GB | 4.3 GB |
GPUs that can run Gemma 4 E2B
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 | ~46 tok/s |
| RTX 4060 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~56 tok/s |
| RTX 4060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~59 tok/s |
| Radeon RX 7600 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~50 tok/s |
| RTX 3060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~90 tok/s |
| RTX 3070 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~90 tok/s |
| RTX 5060 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~90 tok/s |
| RTX 5060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~90 tok/s |
| RTX 3070 Ti 8GB | 8.0 GB | Fits | Q8_0 | 128K | ~121 tok/s |
| Intel Arc B570 10GB | 10.0 GB | Fits | Q8_0 | 128K | ~60 tok/s |
| RTX 3080 10GB | 10.0 GB | Fits | Q8_0 | 128K | ~150 tok/s |
| Apple M4 16GB | 10.7 GB | Fits | Q8_0 | 128K | ~25 tok/s |
| RTX 2080 Ti 11GB | 11.0 GB | Fits | Q8_0 | 128K | ~123 tok/s |
| RTX 2060 12GB | 12.0 GB | Fits | FP16 / BF16 | 128K | ~68 tok/s |
About Gemma 4 E2B
An on-device model with about 2.3B effective parameters. The file holds 5.1B parameters because each layer has its own small per-layer embedding table; runtimes can keep those tables in system memory, so real GPU use can be lower than shown here. It accepts text, images and audio.
| Spec | Value |
|---|---|
| Parameters | 5.1B |
| Layers | 35 |
| Hidden size | 1536 |
| Attention heads / KV heads | 8 / 1 (varies by layer) |
| Head dimension | 256 |
| Vocabulary | 262,144 |
| Max context (used here) | 128K tokens |
| Layers with a growing cache | 3 of 35 |
| KV cache per token (FP16) | 6 KB |
| Released | 2026-04 |
| License | Apache 2.0 |
Gemma 4 E2B VRAM: common questions
Why is Gemma 4 E2B bigger than 2B?
The E stands for effective. It has 2.3B parameters in the compute path plus per-layer embedding tables, for 5.1B in total. The tables are cheap lookups, so speed is like a 2B model.
How much VRAM does Gemma 4 E2B need?
About 4.0 GB at Q4_K_M with an 8K context, 6.1 GB at Q8_0 and 10.6 GB at FP16. Longer contexts add 0.01 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Gemma 4 E2B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the RTX 3050 8GB (8.0 GB usable).