Skip to content

Gemma 3 27B IT VRAM requirements

How much VRAM Gemma 3 27B IT needs
4-bit weights: 15.5 GBWith 8K context: 17.6 GBWith 32K context: 19.5 GB8-bit, 8K context: 29.3 GB
Parameters
27.4B
Longest context
128K
Released
Mar 2025
License
Gemma
Access
Gated on the Hub
Strongest at
Text and images

Hugging Face asks you to accept its terms before it serves the weights. Its licence (Gemma) is not a standard open-source one: check the terms before commercial use.

How much VRAM Gemma 3 27B IT needs

Memory Gemma 3 27B IT needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit51.1 GB53.2 GB55.1 GB62.6 GB
8-bit27.1 GB29.3 GB31.2 GB38.7 GB
6-bit20.9 GB23.1 GB25.0 GB32.5 GB
5-bit18.2 GB20.3 GB22.2 GB29.7 GB
4-bit15.5 GB17.6 GB19.5 GB27.0 GB
3-bit12.5 GB14.6 GB16.5 GB24.0 GB
2-bit9.6 GB11.7 GB13.6 GB21.1 GB
1.6-bit5.3 GB7.4 GB9.3 GB16.8 GB

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

GPUs and Macs that run Gemma 3 27B IT

The best way Gemma 3 27B IT runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Gemma 3 27B IT on RTX 4060 report
RTX 5060Too bigPDF: Gemma 3 27B IT on RTX 5060 report
RTX 3070Too bigPDF: Gemma 3 27B IT on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Gemma 3 27B IT on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Gemma 3 27B IT on RTX 3060 12 GB report
Arc B580Too bigPDF: Gemma 3 27B IT on Arc B580 report
RTX 4070Too bigPDF: Gemma 3 27B IT on RTX 4070 report
RTX 4070 SuperToo bigPDF: Gemma 3 27B IT on RTX 4070 Super report
RTX 5070Too bigPDF: Gemma 3 27B IT on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Gemma 3 27B IT on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Gemma 3 27B IT on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Gemma 3 27B IT on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Gemma 3 27B IT on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Gemma 3 27B IT on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Gemma 3 27B IT on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Gemma 3 27B IT on RTX 5070 Ti report
RTX 5080Too bigPDF: Gemma 3 27B IT on RTX 5080 report
Radeon RX 7900 XTX5-bit36K28 tok/s2.4 sPDF: Gemma 3 27B IT on Radeon RX 7900 XTX report
RTX 30905-bit36K33 tok/s1.5 sPDF: Gemma 3 27B IT on RTX 3090 report
RTX 40905-bit36K36 tok/s0.7 sPDF: Gemma 3 27B IT on RTX 4090 report
Radeon AI PRO R97008-bit20K13 tok/s1.5 sPDF: Gemma 3 27B IT on Radeon AI PRO R9700 report
RTX 50908-bit20K42 tok/s0.5 sPDF: Gemma 3 27B IT on RTX 5090 report
2× RTX 30908-bit128K23 tok/s0.9 sPDF: Gemma 3 27B IT on 2× RTX 3090 report
2× RTX 40908-bit128K24 tok/s0.4 sPDF: Gemma 3 27B IT on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K32 tok/s0.8 sPDF: Gemma 3 27B IT on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K23 tok/s0.6 sPDF: Gemma 3 27B IT on RTX 6000 Ada report
L40S8-bit128K21 tok/s1.0 sPDF: Gemma 3 27B IT on L40S report
2× RTX 50908-bit128K42 tok/s0.3 sPDF: Gemma 3 27B IT on 2× RTX 5090 report
4× RTX 30908-bit128K23 tok/s0.5 sPDF: Gemma 3 27B IT on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K42 tok/s0.4 sPDF: Gemma 3 27B IT on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Gemma 3 27B IT on Mac mini M4 16 GB report
Mac mini M4 24 GB3-bit16K7.5 tok/s35 sPDF: Gemma 3 27B IT on Mac mini M4 24 GB report
Mac mini M4 32 GB4-bit44K6.1 tok/s35 sPDF: Gemma 3 27B IT on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB3-bit16K16 tok/s17 sPDF: Gemma 3 27B IT on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit72K7.8 tok/s17 sPDF: Gemma 3 27B IT on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K7.8 tok/s17 sPDF: Gemma 3 27B IT on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K14 tok/s8.7 sPDF: Gemma 3 27B IT on M4 Max 128 GB report
M5 Max 128 GB8-bit128K15 tok/s2.5 sPDF: Gemma 3 27B IT on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K18 tok/s4.9 sPDF: Gemma 3 27B IT on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K18 tok/s4.9 sPDF: Gemma 3 27B IT on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K18 tok/s4.9 sPDF: Gemma 3 27B IT on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K25 tok/s1.5 sPDF: Gemma 3 27B IT on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K5.5 tok/s5.0 sPDF: Gemma 3 27B IT on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K5.5 tok/s5.0 sPDF: Gemma 3 27B IT on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K7.5 tok/s1.7 sPDF: Gemma 3 27B IT on RTX Spark report
DGX Spark8-bit128K6.8 tok/s1.7 sPDF: Gemma 3 27B IT on DGX Spark report
Dual DGX Spark8-bit128K6.8 tok/s1.0 sPDF: Gemma 3 27B IT on Dual DGX Spark report
DGX Station GB3008-bit128K135 tok/s0.1 sPDF: Gemma 3 27B IT on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K42 tok/s0.2 sPDF: Gemma 3 27B IT on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K42 tok/s0.1 sPDF: Gemma 3 27B IT on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K42 tok/s0.1 sPDF: Gemma 3 27B IT on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K47 tok/s1.1 sPDF: Gemma 3 27B IT on A100 80 GB SXM report
H100 SXM8-bit128K73 tok/s0.3 sPDF: Gemma 3 27B IT on H100 SXM report
4× H200 NVL8-bit128K99 tok/s0.1 sPDF: Gemma 3 27B IT on 4× H200 NVL report
8× H200 NVL8-bit128K99 tok/s0.1 sPDF: Gemma 3 27B IT 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

llama.cpp
llama-server -hf unsloth/gemma-3-27b-it-GGUF
vLLM
vllm serve google/gemma-3-27b-it

Questions

How much VRAM does Gemma 3 27B IT need?

Gemma 3 27B IT needs 15.5 GB for its weights at 4-bit, 17.6 GB with an 8K context and 19.5 GB with 32K. At 8-bit, 29.3 GB.

What is the minimum VRAM for Gemma 3 27B IT?

About 14.6 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 Radeon RX 7900 XTX.

How big is Gemma 3 27B IT?

27.4B parameters: 51.1 GB of weights at 16-bit, 15.5 GB at 4-bit.

Can Gemma 3 27B IT run on an RTX 4090?

Yes. At 5-bit it holds 36K of context and writes about 36 tokens a second.

More models