Skip to content

Gemma 4 E4B VRAM requirements

How much VRAM Gemma 4 E4B needs
4-bit weights: 4.5 GBWith 8K context: 5.7 GBWith 32K context: 6.1 GB8-bit, 8K context: 9.1 GB
Parameters
8B
Active per token
4.5B
Longest context
128K
Released
Mar 2026
License
Apache 2.0
Strongest at
Chat, Coding, Docs

How much VRAM Gemma 4 E4B needs

Memory Gemma 4 E4B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit14.9 GB16.1 GB16.4 GB17.9 GB
8-bit7.9 GB9.1 GB9.5 GB11.0 GB
6-bit6.1 GB7.3 GB7.7 GB9.2 GB
5-bit5.3 GB6.5 GB6.8 GB8.3 GB
4-bit4.5 GB5.7 GB6.1 GB7.6 GB
3-bit3.6 GB4.8 GB5.2 GB6.7 GB
2-bit2.8 GB4.0 GB4.3 GB5.8 GB
1.6-bit1.5 GB2.7 GB3.1 GB4.6 GB

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

GPUs and Macs that run Gemma 4 E4B

The best way Gemma 4 E4B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40606-bit28K42 tok/s0.6 sPDF: Gemma 4 E4B on RTX 4060 report
RTX 50606-bit28K67 tok/s0.4 sPDF: Gemma 4 E4B on RTX 5060 report
RTX 30706-bit28K67 tok/s0.4 sPDF: Gemma 4 E4B on RTX 3070 report
RTX 3080 10 GB8-bit32K86 tok/s0.3 sPDF: Gemma 4 E4B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit128K43 tok/s0.7 sPDF: Gemma 4 E4B on RTX 3060 12 GB report
Arc B5808-bit128K37 tok/s0.5 sPDF: Gemma 4 E4B on Arc B580 report
RTX 40708-bit128K59 tok/s0.3 sPDF: Gemma 4 E4B on RTX 4070 report
RTX 4070 Super8-bit128K59 tok/s0.2 sPDF: Gemma 4 E4B on RTX 4070 Super report
RTX 50708-bit128K77 tok/s0.3 sPDF: Gemma 4 E4B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit128K35 tok/s0.4 sPDF: Gemma 4 E4B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit128K53 tok/s0.4 sPDF: Gemma 4 E4B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit128K45 tok/s0.4 sPDF: Gemma 4 E4B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit128K61 tok/s0.2 sPDF: Gemma 4 E4B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit128K77 tok/s0.2 sPDF: Gemma 4 E4B on RTX 4070 Ti Super report
RTX 4080 Super8-bit128K84 tok/s0.2 sPDF: Gemma 4 E4B on RTX 4080 Super report
RTX 5070 Ti8-bit128K100 tok/s0.2 sPDF: Gemma 4 E4B on RTX 5070 Ti report
RTX 50808-bit128K106 tok/s0.2 sPDF: Gemma 4 E4B on RTX 5080 report
Radeon RX 7900 XTX8-bit128K86 tok/s0.4 sPDF: Gemma 4 E4B on Radeon RX 7900 XTX report
RTX 30908-bit128K103 tok/s0.2 sPDF: Gemma 4 E4B on RTX 3090 report
RTX 40908-bit128K110 tok/s0.1 sPDF: Gemma 4 E4B on RTX 4090 report
Radeon AI PRO R97008-bit128K61 tok/s0.2 sPDF: Gemma 4 E4B on Radeon AI PRO R9700 report
RTX 50908-bit128K177 tok/s0.1 sPDF: Gemma 4 E4B on RTX 5090 report
2× RTX 30908-bit128K103 tok/s0.1 sPDF: Gemma 4 E4B on 2× RTX 3090 report
2× RTX 40908-bit128K110 tok/s0.1 sPDF: Gemma 4 E4B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K141 tok/s0.1 sPDF: Gemma 4 E4B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K106 tok/s0.1 sPDF: Gemma 4 E4B on RTX 6000 Ada report
L40S8-bit128K96 tok/s0.1 sPDF: Gemma 4 E4B on L40S report
2× RTX 50908-bit128K177 tok/s0.1 sPDF: Gemma 4 E4B on 2× RTX 5090 report
4× RTX 30908-bit128K103 tok/s0.1 sPDF: Gemma 4 E4B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K177 tok/s0.1 sPDF: Gemma 4 E4B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit80K17 tok/s5.4 sPDF: Gemma 4 E4B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K17 tok/s5.4 sPDF: Gemma 4 E4B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit128K17 tok/s5.4 sPDF: Gemma 4 E4B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K36 tok/s2.7 sPDF: Gemma 4 E4B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K36 tok/s2.7 sPDF: Gemma 4 E4B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K36 tok/s2.7 sPDF: Gemma 4 E4B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K59 tok/s1.4 sPDF: Gemma 4 E4B on M4 Max 128 GB report
M5 Max 128 GB8-bit128K65 tok/s0.4 sPDF: Gemma 4 E4B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K73 tok/s0.8 sPDF: Gemma 4 E4B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K73 tok/s0.8 sPDF: Gemma 4 E4B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K73 tok/s0.8 sPDF: Gemma 4 E4B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K96 tok/s0.2 sPDF: Gemma 4 E4B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K26 tok/s0.8 sPDF: Gemma 4 E4B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K26 tok/s0.8 sPDF: Gemma 4 E4B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K36 tok/s0.3 sPDF: Gemma 4 E4B on RTX Spark report
DGX Spark8-bit128K33 tok/s0.3 sPDF: Gemma 4 E4B on DGX Spark report
Dual DGX Spark8-bit128K33 tok/s0.2 sPDF: Gemma 4 E4B on Dual DGX Spark report
DGX Station GB3008-bit128K422 tok/s0.1 sPDF: Gemma 4 E4B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K177 tok/s0.1 sPDF: Gemma 4 E4B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K177 tok/s0.1 sPDF: Gemma 4 E4B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K177 tok/s0.1 sPDF: Gemma 4 E4B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K195 tok/s0.2 sPDF: Gemma 4 E4B on A100 80 GB SXM report
H100 SXM8-bit128K277 tok/s0.1 sPDF: Gemma 4 E4B on H100 SXM report
4× H200 NVL8-bit128K345 tok/s0.1 sPDF: Gemma 4 E4B on 4× H200 NVL report
8× H200 NVL8-bit128K345 tok/s0.1 sPDF: Gemma 4 E4B 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:e4b
llama.cpp
llama-server -hf unsloth/gemma-4-E4B-it-GGUF
vLLM
vllm serve google/gemma-4-E4B-it
MLX (Mac)
mlx_lm.chat --model mlx-community/gemma-4-e4b-it-4bit

Questions

How much VRAM does Gemma 4 E4B need?

Gemma 4 E4B needs 4.5 GB for its weights at 4-bit, 5.7 GB with an 8K context and 6.1 GB with 32K. At 8-bit, 9.1 GB.

What is the minimum VRAM for Gemma 4 E4B?

About 4.8 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 E4B?

8B parameters (4.5B active per token): 14.9 GB of weights at 16-bit, 4.5 GB at 4-bit.

Can Gemma 4 E4B run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 110 tokens a second.

More models