What LLM can I run on 8 GB of RAM?

With 8 GB you can comfortably run 4-bit quantized models up to about 7–8B parameters, like Llama 3.1 8B or DeepSeek R1 7B. Smaller models such as Llama 3.2 3B or Gemma 3 4B run even faster and leave room for other apps.

Do I need a GPU to run a local LLM?

No. Tools like Ollama, LM Studio and llama.cpp run models on the CPU. A GPU (or Apple Silicon unified memory) makes generation several times faster, but RAM is the real gatekeeper — the model has to fit in memory.

What is quantization (Q4, Q8)?

Quantization shrinks a model by storing its weights with fewer bits. Q4 (4-bit) cuts the size roughly 4× compared to the original 16-bit weights with only a small quality loss — it's the standard way to run LLMs on consumer hardware.

How accurate is this test?

Browsers cap some signals for privacy: reported memory maxes out at 8 GB and GPU names are generalized. Treat the result as a solid estimate, and set your real RAM manually for a precise verdict. Everything runs in your browser — no data is stored.

Which apps can I use to run models locally?

The most popular are Ollama, LM Studio, llama.cpp and Jan. Pick a model this test marks as green, download its Q4 build, and you're chatting offline in minutes.

LOCAL AI CHECK

What LLM can I run?

One click, zero downloads. We read your browser's hardware signals — GPU, memory, CPU cores, WebGPU — and show which local AI models your machine can actually handle. Nothing leaves your device.

Your hardware

Reading your hardware signals…

Verdict

Your machine can run Qwen 2.5 VL 7B 🚀

Every model marked “Runs” below fits comfortably in 8 GB of memory. Grab Ollama or LM Studio and you're chatting offline in minutes.

Which models fit in 8 GB?

73 / 73 models

Model	Params	Download (Q4)	Min RAM	Verdict
Qwen 3 0.6BAlibaba	0.6B	0.4 GB	2 GB	Runs
Gemma 3 1BGoogle	1B	0.6 GB	3 GB	Runs
Llama 3.2 1BMeta	1.2B	0.7 GB	3 GB	Runs
Qwen 3 1.7BAlibaba	1.7B	1.0 GB	3 GB	Runs
SmolLM2 1.7BHugging Face	1.7B	1.0 GB	3 GB	Runs
DeepSeek R1 1.5BDeepSeek	1.8B	1.1 GB	3 GB	Runs
DeepSeek-OCRDeepSeek	3B (A0.57B)	1.8 GB	4 GB	Runs
Ministral 3 3BMistral AI	3B	1.8 GB	4 GB	Runs
Llama 3.2 3BMeta	3.2B	1.9 GB	4 GB	Runs
Phi-4 Mini 3.8BMicrosoft	3.8B	2.3 GB	6 GB	Runs
Qwen 3 4BAlibaba	4B	2.4 GB	6 GB	Runs
Qwen 3.5 4BAlibaba	4B	2.4 GB	6 GB	Runs
Gemma 3 4BGoogle	4.3B	2.6 GB	6 GB	Runs
Gemma 4 E2BGoogle	5.1B (A2.3B)	3.1 GB	6 GB	Runs
Mistral 7BMistral AI	7.2B	4.4 GB	8 GB	Runs
Qwen 2.5 Coder 7BAlibaba	7.6B	4.6 GB	8 GB	Runs
DeepSeek R1 7BDeepSeek	7.6B	4.6 GB	8 GB	Runs
Gemma 3n E4BGoogle	7.8B (A4B)	4.7 GB	8 GB	Runs
Llama 3.1 8BMeta	8B	4.9 GB	8 GB	Runs
DeepSeek R1 8BDeepSeek	8B	4.9 GB	8 GB	Runs
Gemma 4 E4BGoogle	8B (A4.5B)	4.9 GB	8 GB	Runs
Qwen3-VL 8BAlibaba	8B	4.9 GB	8 GB	Runs
Ministral 3 8BMistral AI	8B	4.9 GB	8 GB	Runs
Qwen 3 8BAlibaba	8.2B	5.0 GB	8 GB	Runs
Granite 3.3 8BIBM	8.2B	5.0 GB	8 GB	Runs
Qwen 2.5 VL 7BAlibaba	8.3B	5.0 GB	8 GB	Runs
Qwen 3.5 9BAlibaba	9B	5.5 GB	12 GB	Too heavy
GLM-4.6V-FlashZ.ai	9B	5.5 GB	12 GB	Too heavy
Gemma 4 12BGoogle	12B	7.3 GB	12 GB	Too heavy
Mellum 2 12B-A2.5BJetBrains	12B (A2.5B)	7.3 GB	12 GB	Too heavy
Gemma 3 12BGoogle	12.2B	7.4 GB	12 GB	Too heavy
Mistral Nemo 12BMistral AI	12.2B	7.4 GB	12 GB	Too heavy
OLMo 2 13BAi2	13.7B	8.3 GB	12 GB	Too heavy
Ministral 3 14BMistral AI	14B	8.5 GB	16 GB	Too heavy
Phi-4 14BMicrosoft	14.7B	8.9 GB	16 GB	Too heavy
Qwen 3 14BAlibaba	14.8B	9.0 GB	16 GB	Too heavy
DeepSeek R1 14BDeepSeek	14.8B	9.0 GB	16 GB	Too heavy
Phi-4 Reasoning Vision 15BMicrosoft	15B	9.1 GB	16 GB	Too heavy
GPT-OSS 20BOpenAI	20.9B (A3.6B)	12.7 GB	24 GB	Too heavy
Codestral 22BMistral AI	22.2B	13.5 GB	24 GB	Too heavy
Mistral Small 3.1 24BMistral AI	24B	14.6 GB	24 GB	Too heavy
Devstral 24BMistral AI	24B	14.6 GB	24 GB	Too heavy
Magistral Small 1.2Mistral AI	24B	14.6 GB	24 GB	Too heavy
Devstral Small 2 24BMistral AI	24B	14.6 GB	24 GB	Too heavy
Gemma 4 26B A4BGoogle	25.2B (A3.8B)	15.3 GB	24 GB	Too heavy
Qwen 3.5 27BAlibaba	27B	16.4 GB	24 GB	Too heavy
Qwen 3.6 27BAlibaba	27B	16.4 GB	24 GB	Too heavy
Gemma 3 27BGoogle	27.4B	16.6 GB	24 GB	Too heavy
Qwen3-VL 30B-A3BAlibaba	30B (A3B)	18.2 GB	32 GB	Too heavy
Qwen 3 30B-A3BAlibaba	30.5B (A3.3B)	18.5 GB	32 GB	Too heavy
Gemma 4 31BGoogle	30.7B	18.6 GB	32 GB	Too heavy
Nemotron 3 Nano 30B-A3BNVIDIA	31.6B (A3.6B)	19.2 GB	32 GB	Too heavy
Granite 4.0 H SmallIBM	32B (A9B)	19.4 GB	32 GB	Too heavy
Qwen 3 32BAlibaba	32.8B	19.9 GB	32 GB	Too heavy
Qwen 2.5 Coder 32BAlibaba	32.8B	19.9 GB	32 GB	Too heavy
QwQ 32BAlibaba	32.8B	19.9 GB	32 GB	Too heavy
DeepSeek R1 32BDeepSeek	32.8B	19.9 GB	32 GB	Too heavy
Qwen3-VL 32BAlibaba	33B	20.0 GB	32 GB	Too heavy
EXAONE 4.5 33BLG AI Research	33B	20.0 GB	32 GB	Too heavy
Qwen 3.5 35B-A3BAlibaba	35B (A3B)	21.2 GB	32 GB	Too heavy
Qwen 3.6 35B-A3BAlibaba	35B (A3B)	21.2 GB	32 GB	Too heavy
Command R 35BCohere	35B	21.2 GB	32 GB	Too heavy
Llama 3.1 70BMeta	70.6B	42.8 GB	64 GB	Too heavy
Llama 3.3 70BMeta	70.6B	42.8 GB	64 GB	Too heavy
DeepSeek R1 70BDeepSeek	70.6B	42.8 GB	64 GB	Too heavy
Qwen3-Next 80B-A3BAlibaba	80B (A3B)	48.5 GB	64 GB	Too heavy
Qwen3 Coder Next 80B-A3BAlibaba	80B (A3B)	48.5 GB	64 GB	Too heavy
Llama 4 ScoutMeta	109B (A17B)	66.1 GB	96 GB	Too heavy
GPT-OSS 120BOpenAI	116.8B (A5.1B)	70.8 GB	96 GB	Too heavy
Mistral Small 4 119BMistral AI	119B (A6.5B)	72.1 GB	96 GB	Too heavy
Nemotron 3 Super 120B-A12BNVIDIA	120B (A12B)	72.8 GB	96 GB	Too heavy
Qwen 3.5 122B-A10BAlibaba	122B (A10B)	74.0 GB	96 GB	Too heavy
Devstral 2 123BMistral AI	123B	74.6 GB	96 GB	Too heavy

Sizes are 4-bit (Q4_K_M) GGUF builds — the standard for running models locally. · Data updated: 2026-06-11

Everything runs in your browser. Nothing is uploaded, nothing is stored. Zero logs.

Best local LLMs by RAM

Best LLMs for 8 GB RAM

26 models run comfortably

What LLM can I run?

Frequently asked questions