Models / gemma-3

gemma-3 by google

5 models (270M to 27B) · ↓ 6.1M downloads a month · from $0.08/h on GTX 1660 SUPER right now. Each model's versions do the same job; quantized ones (AWQ, FP8, 4-bit) need less GPU memory and run on cheaper machines.

gemma-3 270M

Version	Runtime	Weights	GPU memory	Cheapest machine now
Originalgoogle/gemma-3-270mLicence to accept	vLLM	540 MB	~2.9 GB	$0.10/h RTX 3060
itgoogle/gemma-3-270m-itLicence to accept	vLLM	540 MB	~2.9 GB	$0.10/h RTX 3060
it unsloth bnb 4bit (unsloth)unsloth/gemma-3-270m-it-unsloth-bnb-4bit	vLLM	390 MB	2.8 GB	$0.10/h RTX 3060

gemma-3 1B

Version	Runtime	Weights	GPU memory	Cheapest machine now
Originalgoogle/gemma-3-1b-ptLicence to accept	vLLM	2.0 GB	~4.6 GB	$0.10/h RTX 3060
gemma 3 1b itgoogle/gemma-3-1b-itLicence to accept	vLLM	2.0 GB	~4.6 GB	$0.10/h RTX 3060
gemma 3 1b it unsloth bnb 4bit (unsloth)unsloth/gemma-3-1b-it-unsloth-bnb-4bit	vLLM	1.0 GB	3.6 GB	$0.10/h RTX 3060
gemma 3 1b instruct onnx genai int4 emb int8 (Arm)Arm/gemma-3-1b-instruct-onnx-genai-int4-emb-int8	ONNX Runtime GenAI	910 MB	2.7 GB	$0.08/h GTX 1660 SUPER
gemma 3 1b base onnx genai int4 emb int8 (Arm)Arm/gemma-3-1b-base-onnx-genai-int4-emb-int8	ONNX Runtime GenAI	900 MB	2.6 GB	$0.08/h GTX 1660 SUPER

gemma-3 4B

Version	Runtime	Weights	GPU memory	Cheapest machine now
Originalgoogle/gemma-3-4b-ptLicence to accept	vLLM	8.6 GB	~12.5 GB	$0.14/h RTX A4000
gemma 3 4b itgoogle/gemma-3-4b-itLicence to accept	vLLM	8.6 GB	~12.5 GB	$0.14/h RTX A4000

gemma-3 12B

Version	Runtime	Weights	GPU memory	Cheapest machine now
Originalgoogle/gemma-3-12b-ptLicence to accept	vLLM	24 GB	~31.5 GB	$0.41/h RTX PRO 4500
gemma 3 12b itgoogle/gemma-3-12b-itLicence to accept	vLLM	24 GB	~31.5 GB	$0.41/h RTX PRO 4500
gemma 3 12b it quantized W4A16 (abhishekchohan)abhishekchohan/gemma-3-12b-it-quantized-W4A16	vLLM	8.4 GB	15.1 GB	$0.14/h RTX A4000
gemma 3 12b it int4 awq (gaunernst)gaunernst/gemma-3-12b-it-int4-awq	vLLM	9.1 GB	15.7 GB	$0.14/h RTX A4000
gemma 3 12b it heretic v2 (DreamFast)DreamFast/gemma-3-12b-it-heretic-v2	vLLM	24 GB	32.9 GB	$0.45/h Nvidia A40
gemma 3 12b it heretic (DreamFast)DreamFast/gemma-3-12b-it-heretic	vLLM	24 GB	32.9 GB

gemma-3 27B

Version	Runtime	Weights	GPU memory	Cheapest machine now
gemma 3 27b itgoogle/gemma-3-27b-itLicence to accept	vLLM	55 GB	~68.1 GB	$0.89/h 2x Nvidia A40
medgemma 27b itgoogle/medgemma-27b-itLicence to accept	vLLM	55 GB	~68.1 GB	$0.89/h 2x Nvidia A40
medgemma 27b text itgoogle/medgemma-27b-text-itLicence to accept	vLLM	54 GB	~67 GB	$0.89/h 2x Nvidia A40
gemma 3 27b it int4 awq (gaunernst)gaunernst/gemma-3-27b-it-int4-awq	vLLM	18 GB	27.4 GB	$0.34/h 2x RTX 5060 Ti
gemma 3 27b it GPTQ 4b 128g (ISTA-DASLab)ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g	vLLM	17 GB	25.6 GB	$0.34/h 2x RTX 5060 Ti
gemma 3 27b it quantized.w4a16 (RedHatAI)RedHatAI/gemma-3-27b-it-quantized.w4a16	vLLM	20 GB

Model details come from Hugging Face and are refreshed daily; prices follow the live offers. GPU memory is GridPup's estimate: the weights, room for an 8,192-token conversation (chat models) and the runtime's own memory. Large models are split across the GPUs of one machine ("4x RTX 4090").

Loading…