← CanIRunLocal · All hardware · All models

Can I run Gemma 3 27B locally?

Google · 27.4B parameters · chat, vision.

Ollama weight files: Q4_K_M 17 GB (gemma3:27b); Q8_0 30 GB (gemma3:27b-it-q8_0).

Yes on 48 machines in this list. The smallest is RTX 3080 10GB: Q4_K_M, about 19 GB, roughly 4 tokens per second, via Ollama. ollama run gemma3:27b

DGX Spark 128GB: Q8_0, ~32 GB, fits in fast memory, ~7 tok/s — ollama run gemma3:27b-it-q8_0
Mac Studio M5 Max 128GB: Q8_0, ~32 GB, fits in fast memory, ~16 tok/s — ollama run gemma3:27b-it-q8_0
Mac Studio M5 Ultra 256GB: Q8_0, ~32 GB, fits in fast memory, ~32 tok/s — ollama run gemma3:27b-it-q8_0

Setup on DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte)

Ollama: ollama run gemma3:27b-it-q8_0
LM Studio: lms get unsloth/gemma-3-27b-it-GGUF@Q8_0
llama.cpp: llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 -c 8192
vLLM: vllm serve google/gemma-3-27b-it --max-model-len 8192 --gpu-memory-utilization 0.9

Open Gemma 3 27B in the checker

Hardware that fits

Hardware	Runtime	Quant	Memory	Fit	Speed	Command
RTX 3080 10GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~4 tok/s	`ollama run gemma3:27b`
RTX 3060 12GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~5 tok/s	`ollama run gemma3:27b`
RTX 4070 12GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~5 tok/s	`ollama run gemma3:27b`
RTX 5070 12GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~5 tok/s	`ollama run gemma3:27b`
RTX 4060 Ti 16GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~6 tok/s	`ollama run gemma3:27b`
RTX 4070 Ti Super 16GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~9 tok/s	`ollama run gemma3:27b`
RTX 4080 16GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~9 tok/s	`ollama run gemma3:27b`
RTX 5080 16GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~10 tok/s	`ollama run gemma3:27b`
RX 7800 XT 16GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~9 tok/s	`ollama run gemma3:27b`
RX 9070 XT 16GB	Ollama	Q4_K_M	19 GB	partial offload (slower)	~9 tok/s	`ollama run gemma3:27b`
RTX 3090 24GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~33 tok/s	`ollama run gemma3:27b`
RTX 4090 24GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~36 tok/s	`ollama run gemma3:27b`
RX 7900 XTX 24GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~34 tok/s	`ollama run gemma3:27b`
M1 Pro 32GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~9 tok/s	`ollama run gemma3:27b`
M2 Pro 32GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~9 tok/s	`ollama run gemma3:27b`
M4 32GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~6 tok/s	`ollama run gemma3:27b`
M3 Pro 36GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~7 tok/s	`ollama run gemma3:27b`
Mac Studio M4 Max (32c GPU) 36GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~19 tok/s	`ollama run gemma3:27b`
Mac Studio M5 Max (32c GPU) 36GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~22 tok/s	`ollama run gemma3:27b`
RTX 5090 32GB	Ollama	Q4_K_M	19 GB	fits in fast memory	~63 tok/s	`ollama run gemma3:27b`
M4 Pro 48GB	Ollama	Q8_0	32 GB	fits in fast memory	~7 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M4 Max (40c GPU) 48GB	Ollama	Q8_0	32 GB	fits in fast memory	~15 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M5 Max (40c GPU) 48GB	Ollama	Q8_0	32 GB	fits in fast memory	~16 tok/s	`ollama run gemma3:27b-it-q8_0`
M1 Max 64GB	Ollama	Q8_0	32 GB	fits in fast memory	~11 tok/s	`ollama run gemma3:27b-it-q8_0`
M2 Max 64GB	Ollama	Q8_0	32 GB	fits in fast memory	~11 tok/s	`ollama run gemma3:27b-it-q8_0`
M3 Max 64GB	Ollama	Q8_0	32 GB	fits in fast memory	~11 tok/s	`ollama run gemma3:27b-it-q8_0`
M4 Max 64GB	Ollama	Q8_0	32 GB	fits in fast memory	~15 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M4 Max (40c GPU) 64GB	Ollama	Q8_0	32 GB	fits in fast memory	~15 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M5 Max (40c GPU) 64GB	Ollama	Q8_0	32 GB	fits in fast memory	~16 tok/s	`ollama run gemma3:27b-it-q8_0`
DGX Spark 64GB (ASUS, Dell, HP, MSI, Acer, Gigabyte)	Ollama	Q8_0	32 GB	fits in fast memory	~7 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M3 Ultra 96GB	Ollama	Q8_0	32 GB	fits in fast memory	~22 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M5 Ultra 96GB	Ollama	Q8_0	32 GB	fits in fast memory	~32 tok/s	`ollama run gemma3:27b-it-q8_0`
M3 Max 128GB	Ollama	Q8_0	32 GB	fits in fast memory	~11 tok/s	`ollama run gemma3:27b-it-q8_0`
M4 Max 128GB	Ollama	Q8_0	32 GB	fits in fast memory	~15 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M4 Max (40c GPU) 128GB	Ollama	Q8_0	32 GB	fits in fast memory	~15 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M5 Max (40c GPU) 128GB	Ollama	Q8_0	32 GB	fits in fast memory	~16 tok/s	`ollama run gemma3:27b-it-q8_0`
DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte)	Ollama	Q8_0	32 GB	fits in fast memory	~7 tok/s	`ollama run gemma3:27b-it-q8_0`
M2 Ultra 192GB	Ollama	Q8_0	32 GB	fits in fast memory	~21 tok/s	`ollama run gemma3:27b-it-q8_0`
2x Mac Studio M5 Max (40c GPU) 128GB	llama.cpp	Q8_0 GGUF	30.7 GB	fits in fast memory	~15 tok/s	`llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 --rpc node2:50052 -c 8192 # run rpc-server on each other node`
Mac Studio M3 Ultra 256GB	Ollama	Q8_0	32 GB	fits in fast memory	~22 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M5 Ultra (64c GPU) 256GB	Ollama	Q8_0	32 GB	fits in fast memory	~32 tok/s	`ollama run gemma3:27b-it-q8_0`
Mac Studio M5 Ultra (80c GPU) 256GB	Ollama	Q8_0	32 GB	fits in fast memory	~32 tok/s	`ollama run gemma3:27b-it-q8_0`
2x DGX Spark 128GB	llama.cpp	Q8_0 GGUF	30.7 GB	fits in fast memory	~6 tok/s	`llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 --rpc node2:50052 -c 8192 # run rpc-server on each other node`
4x Mac Studio M5 Max (40c GPU) 128GB	llama.cpp	Q8_0 GGUF	30.7 GB	fits in fast memory	~13 tok/s	`llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 --rpc node2:50052,node3:50052,node4:50052 -c 8192 # run rpc-server on each other node`
2x Mac Studio M5 Ultra (80c GPU) 256GB	llama.cpp	Q8_0 GGUF	30.7 GB	fits in fast memory	~28 tok/s	`llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 --rpc node2:50052 -c 8192 # run rpc-server on each other node`
Mac Studio M3 Ultra (80c GPU) 512GB	Ollama	Q8_0	32 GB	fits in fast memory	~22 tok/s	`ollama run gemma3:27b-it-q8_0`
4x DGX Spark 128GB	llama.cpp	Q8_0 GGUF	30.7 GB	fits in fast memory	~5 tok/s	`llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 --rpc node2:50052,node3:50052,node4:50052 -c 8192 # run rpc-server on each other node`
4x Mac Studio M5 Ultra (80c GPU) 256GB	llama.cpp	Q8_0 GGUF	30.7 GB	fits in fast memory	~25 tok/s	`llama-server -hf unsloth/gemma-3-27b-it-GGUF:Q8_0 --rpc node2:50052,node3:50052,node4:50052 -c 8192 # run rpc-server on each other node`

Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-16.

A project by OnDevice Labs · llms.txt