How many AI models can I run with 48 GB VRAM?

With 48 GB VRAM, you can comfortably run 82 AI models at recommended quantization levels. An additional 8 models fit with tight memory margins.

What is the largest model I can run with 48 GB VRAM?

The largest model you can run with 48 GB VRAM is Qwen 2.5 72B (72B parameters) at Q4_K_M quantization, requiring 44.7 GB.

Best AI Models for 48 GB VRAM

Q: What GPUs have 48 GB VRAM?

GPUs with 48 GB VRAM include: NVIDIA RTX 6000 Ada Generation, NVIDIA RTX A6000, NVIDIA RTX PRO 5000 Blackwell.

Professional — large 70B models at higher quality quants. Here are all 90 models you can run locally with 48 GB of memory.

Hardware with 48 GB Memory

NVIDIA RTX 6000 Ada Generation NVIDIA RTX A6000 NVIDIA RTX PRO 5000 Blackwell Mac mini M4 Pro 48GB MacBook Pro M3 Max 48GB MacBook Pro M4 Max 48GB MacBook Pro M4 Pro 48GB MacBook Pro M5 Max 48GB MacBook Pro M5 Pro 48GB

Runs Comfortably (82)

These models fit with room to spare for context window and OS overhead.

Model	Params	Quantization	VRAM	Quality
Dolphin Mixtral 8x7B	47B	Q4_K_M	26 GB	4
Mixtral 8x7B	47B	Q5_K_M	34.4 GB	4
Command R 35B	35B	Q8_0	37 GB	5
Qwen 3.5 35B A3B	35B	Q8_0	20 GB	5
Nous Hermes 2 34B	34B	Q8_0	36 GB	5
Yi 1.5 34B	34B	Q8_0	36 GB	5
Laguna XS 2.1	33B	Q8_0	39 GB	5
WizardCoder 33B	33B	Q8_0	38.5 GB	5
Aya Expanse 32B	32B	Q8_0	38 GB	5
Cogito 32B	32B	Q8_0	37 GB	4
DeepSeek R1 32B	32B	Q8_0	34 GB	5
Qwen 2.5 32B	32B	Q8_0	34 GB	5
Qwen 2.5 Coder 32B	32B	Q8_0	39 GB	5
Qwen 3 32B	32B	Q8_0	39 GB	5
QwQ 32B	32B	Q8_0	37 GB	4
Gemma 4 31B	31B	Q8_0	38 GB	5
Qwen 3 30B-A3B (MoE)	30B	Q8_0	37 GB	5
Gemma 2 27B	27B	Q8_0	29 GB	5
Gemma 3 27B	27B	Q8_0	34 GB	5
Qwen 3.5 27B	27B	Q8_0	33 GB	5
Qwen 3.6 27B	27B	Q8_0	34 GB	5
Gemma 4 26B	26B	Q8_0	30 GB	5
Devstral 24B	24B	Q8_0	29 GB	5
Magistral Small 24B	24B	Q8_0	28 GB	4
Mistral Small 3.1 24B	24B	Q8_0	30 GB	5
Codestral 22B	22B	Q8_0	24 GB	5
gpt-oss 20B	21B	MXFP4	15 GB	4
StarCoder2 15B	15B	Q8_0	17 GB	5
DeepSeek R1 14B	14B	Q8_0	16 GB	5
Phi-4 14B	14B	Q8_0	16 GB	5
Phi-4 Reasoning 14B	14B	F16	32 GB	5
Qwen 2.5 14B	14B	Q8_0	16 GB	5
Qwen 2.5 Coder 14B	14B	F16	33 GB	5
Qwen 3 14B	14B	F16	33 GB	5
Gemma 3 12B	12B	F16	28 GB	5
Mistral Nemo 12B	12B	F16	28 GB	5
Llama 3.2 Vision 11B	11B	F16	26 GB	5
Falcon 3 10B	10B	F16	24 GB	5
Gemma 2 9B	9B	F16	20 GB	5
Qwen 3.5 9B	9B	Q8_0	11.5 GB	5
Yi 1.5 9B	9B	F16	20 GB	5
Yi Coder 9B	9B	F16	21 GB	5
Aya Expanse 8B	8B	Q8_0	10.5 GB	5
Cogito 8B	8B	F16	19.5 GB	5
DeepSeek R1 8B	8B	F16	20 GB	5
Dolphin 3 8B	8B	F16	18 GB	5
Granite 3.3 8B	8B	F16	18 GB	5
Llama 3.1 8B	8B	F16	18 GB	5
Nemotron 3 Nano 8B	8B	F16	19.5 GB	5
Nous Hermes 2 8B	8B	F16	18 GB	5
Qwen 3 8B	8B	F16	20 GB	5
Codestral Mamba 7B	7B	F16	17 GB	5
DeepSeek R1 7B	7B	F16	16 GB	5
Falcon 3 7B	7B	F16	17.5 GB	5
InternLM 2.5 7B	7B	F16	16 GB	5
Mistral 7B	7B	F16	16 GB	5
OpenChat 3.5 7B	7B	F16	17 GB	5
Qwen 2.5 7B	7B	F16	16 GB	5
Qwen 2.5 Coder 7B	7B	F16	16 GB	5
Qwen 2.5 VL 7B	7B	F16	18.5 GB	5
StarCoder2 7B	7B	F16	16 GB	5
WizardLM 2 7B	7B	F16	17 GB	5
Yi 1.5 6B	6B	F16	14 GB	5
Gemma 3 4B	4B	F16	11.5 GB	5
Gemma 3n E4B	4B	F16	11 GB	5
Gemma 4 E4B	4B	Q8_0	10 GB	5
Qwen 3 4B	4B	F16	11 GB	5
Qwen 3.5 4B	4B	Q8_0	6.5 GB	4
Phi-3 Mini 3.8B	3.8B	F16	9.6 GB	5
Phi-4 Mini 3.8B	3.8B	F16	10.5 GB	5
Llama 3.2 3B	3B	F16	8 GB	5
StarCoder2 3B	3B	F16	8 GB	5
Gemma 2 2B	2B	F16	6 GB	5
Gemma 3n E2B	2B	F16	6.5 GB	5
Gemma 4 E2B	2B	Q8_0	6 GB	5
Qwen 3.5 2B	2B	Q8_0	4.5 GB	4
SmolLM2 1.7B	1.7B	F16	4.4 GB	5
DeepSeek R1 1.5B	1.5B	F16	5 GB	5
Gemma 3 1B	1B	F16	3.5 GB	5
Llama 3.2 1B	1B	F16	4 GB	5
Qwen 3.5 0.8B	0.8B	Q8_0	2 GB	4
Qwen 3 0.6B	0.6B	F16	3.3 GB	5

Tight Fit (8)

These models run but with limited context window. Close other apps to free memory.

Model	Params	Quantization	VRAM	Quality
Qwen 2.5 72B	72B	Q4_K_M	44.7 GB	4
Qwen 2.5 VL 72B	72B	Q4_K_M	41 GB	4
Cogito 70B	70B	Q4_K_M	43 GB	4
DeepSeek R1 70B	70B	Q4_K_M	43.5 GB	4
Llama 3.1 70B	70B	Q4_K_M	43.5 GB	4
Llama 3.3 70B	70B	Q4_K_M	43.5 GB	4
Qwen 3.6 35B-A3B	35B	Q8_0	44 GB	5
InternLM 2.5 20B	20B	F16	42 GB	5

Want to check a specific combination?

Use the compatibility checker to see exactly how a model runs on your specific hardware, with performance estimates.

Open Compatibility Checker

8 GBEntry-level local AI 12 GBThe minimum for a good experience 16 GBThe sweet spot 24 GBSerious local AI