Skip to content

Gemma 4 E2B VRAM requirements

How much VRAM Gemma 4 E2B needs
4-bit weights: 2.9 GBWith 8K context: 3.9 GBWith 32K context: 4.1 GB8-bit, 8K context: 6.1 GB
Parameters
5.1B
Active per token
2.3B
Longest context
128K
Released
Mar 2026
License
Apache 2.0
Strongest at
Chat

How much VRAM Gemma 4 E2B needs

Memory Gemma 4 E2B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit9.5 GB10.6 GB10.7 GB11.3 GB
8-bit5.1 GB6.1 GB6.3 GB6.8 GB
6-bit3.9 GB5.0 GB5.1 GB5.7 GB
5-bit3.4 GB4.4 GB4.6 GB5.1 GB
4-bit2.9 GB3.9 GB4.1 GB4.6 GB
3-bit2.3 GB3.4 GB3.5 GB4.1 GB
2-bit1.8 GB2.8 GB3.0 GB3.5 GB
1.6-bit1.0 GB2.0 GB2.2 GB2.7 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Gemma 4 E2B

The best way Gemma 4 E2B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit128K60 tok/s0.3 sPDF: Gemma 4 E2B on RTX 4060 report
RTX 50608-bit128K96 tok/s0.2 sPDF: Gemma 4 E2B on RTX 5060 report
RTX 30708-bit128K96 tok/s0.2 sPDF: Gemma 4 E2B on RTX 3070 report
RTX 3080 10 GB8-bit128K152 tok/s0.1 sPDF: Gemma 4 E2B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit128K78 tok/s0.3 sPDF: Gemma 4 E2B on RTX 3060 12 GB report
Arc B5808-bit128K67 tok/s0.2 sPDF: Gemma 4 E2B on Arc B580 report
RTX 40708-bit128K106 tok/s0.1 sPDF: Gemma 4 E2B on RTX 4070 report
RTX 4070 Super8-bit128K106 tok/s0.1 sPDF: Gemma 4 E2B on RTX 4070 Super report
RTX 50708-bit128K137 tok/s0.1 sPDF: Gemma 4 E2B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit128K64 tok/s0.2 sPDF: Gemma 4 E2B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit128K96 tok/s0.2 sPDF: Gemma 4 E2B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit128K79 tok/s0.2 sPDF: Gemma 4 E2B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit128K106 tok/s0.1 sPDF: Gemma 4 E2B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit128K137 tok/s0.1 sPDF: Gemma 4 E2B on RTX 4070 Ti Super report
RTX 4080 Super8-bit128K148 tok/s0.1 sPDF: Gemma 4 E2B on RTX 4080 Super report
RTX 5070 Ti8-bit128K174 tok/s0.1 sPDF: Gemma 4 E2B on RTX 5070 Ti report
RTX 50808-bit128K184 tok/s0.1 sPDF: Gemma 4 E2B on RTX 5080 report
Radeon RX 7900 XTX8-bit128K146 tok/s0.2 sPDF: Gemma 4 E2B on Radeon RX 7900 XTX report
RTX 30908-bit128K180 tok/s0.1 sPDF: Gemma 4 E2B on RTX 3090 report
RTX 40908-bit128K191 tok/s0.1 sPDF: Gemma 4 E2B on RTX 4090 report
Radeon AI PRO R97008-bit128K106 tok/s0.1 sPDF: Gemma 4 E2B on Radeon AI PRO R9700 report
RTX 50908-bit128K294 tok/s0.1 sPDF: Gemma 4 E2B on RTX 5090 report
2× RTX 30908-bit128K180 tok/s0.1 sPDF: Gemma 4 E2B on 2× RTX 3090 report
2× RTX 40908-bit128K191 tok/s0.1 sPDF: Gemma 4 E2B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K239 tok/s0.1 sPDF: Gemma 4 E2B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K184 tok/s0.1 sPDF: Gemma 4 E2B on RTX 6000 Ada report
L40S8-bit128K169 tok/s0.1 sPDF: Gemma 4 E2B on L40S report
2× RTX 50908-bit128K294 tok/s0.1 sPDF: Gemma 4 E2B on 2× RTX 5090 report
4× RTX 30908-bit128K180 tok/s0.1 sPDF: Gemma 4 E2B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K294 tok/s0.1 sPDF: Gemma 4 E2B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit128K31 tok/s2.7 sPDF: Gemma 4 E2B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K31 tok/s2.7 sPDF: Gemma 4 E2B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit128K31 tok/s2.7 sPDF: Gemma 4 E2B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K63 tok/s1.3 sPDF: Gemma 4 E2B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K63 tok/s1.3 sPDF: Gemma 4 E2B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K63 tok/s1.3 sPDF: Gemma 4 E2B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K100 tok/s0.7 sPDF: Gemma 4 E2B on M4 Max 128 GB report
M5 Max 128 GB8-bit128K109 tok/s0.2 sPDF: Gemma 4 E2B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K120 tok/s0.4 sPDF: Gemma 4 E2B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K120 tok/s0.4 sPDF: Gemma 4 E2B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K120 tok/s0.4 sPDF: Gemma 4 E2B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K152 tok/s0.1 sPDF: Gemma 4 E2B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K48 tok/s0.4 sPDF: Gemma 4 E2B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K48 tok/s0.4 sPDF: Gemma 4 E2B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K66 tok/s0.1 sPDF: Gemma 4 E2B on RTX Spark report
DGX Spark8-bit128K61 tok/s0.1 sPDF: Gemma 4 E2B on DGX Spark report
Dual DGX Spark8-bit128K61 tok/s0.1 sPDF: Gemma 4 E2B on Dual DGX Spark report
DGX Station GB3008-bit128K609 tok/s0.1 sPDF: Gemma 4 E2B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K294 tok/s0.1 sPDF: Gemma 4 E2B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K294 tok/s0.1 sPDF: Gemma 4 E2B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K294 tok/s0.1 sPDF: Gemma 4 E2B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K321 tok/s0.1 sPDF: Gemma 4 E2B on A100 80 GB SXM report
H100 SXM8-bit128K433 tok/s0.1 sPDF: Gemma 4 E2B on H100 SXM report
4× H200 NVL8-bit128K519 tok/s0.1 sPDF: Gemma 4 E2B on 4× H200 NVL report
8× H200 NVL8-bit128K519 tok/s0.1 sPDF: Gemma 4 E2B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

Ollama
ollama run gemma4:e2b
llama.cpp
llama-server -hf unsloth/gemma-4-E2B-it-GGUF
vLLM
vllm serve google/gemma-4-E2B-it
MLX (Mac)
mlx_lm.chat --model mlx-community/gemma-4-e2b-it-4bit

Questions

How much VRAM does Gemma 4 E2B need?

Gemma 4 E2B needs 2.9 GB for its weights at 4-bit, 3.9 GB with an 8K context and 4.1 GB with 32K. At 8-bit, 6.1 GB.

What is the minimum VRAM for Gemma 4 E2B?

About 3.4 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is Gemma 4 E2B?

5.1B parameters (2.3B active per token): 9.5 GB of weights at 16-bit, 2.9 GB at 4-bit.

Can Gemma 4 E2B run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 191 tokens a second.

More models