Skip to content

NVIDIA GeForce RTX 3090

NVIDIA · 24GB GDDR6X · Can run 50 models

Buy Amazon
Manufacturer NVIDIA
VRAM 24 GB
Memory Type GDDR6X
Architecture Ampere
CUDA Cores 10,496
Tensor Cores 328
Bandwidth 936 GB/s
TDP 350W
MSRP $1,499
Released Sep 24, 2020

AI Notes

The RTX 3090 remains a strong contender for local AI with its generous 24GB of GDDR6X VRAM. It can load 13B models at full precision and run 30B+ models with quantization. Available at steep discounts on the used market, it offers excellent VRAM-per-dollar for AI workloads.

Compatible Models

Model Parameters Best Quant VRAM Used Fit Est. Speed
Qwen 3 0.6B 600M Q4_K_M 2.5 GB Runs ~374 tok/s
Gemma 3 1B 1B Q8_0 2 GB Runs ~468 tok/s
Llama 3.2 1B 1B Q8_0 3 GB Runs ~312 tok/s
DeepSeek R1 1.5B 1.5B Q8_0 3 GB Runs ~312 tok/s
Gemma 2 2B 2B Q8_0 4 GB Runs ~234 tok/s
Gemma 3n E2B 2B Q4_K_M 3.3 GB Runs ~284 tok/s
Llama 3.2 3B 3B Q8_0 5 GB Runs ~187 tok/s
Phi-3 Mini 3.8B 3.8B Q8_0 5.8 GB Runs ~161 tok/s
Phi-4 Mini 3.8B 3.8B Q4_K_M 4.5 GB Runs ~208 tok/s
Gemma 3 4B 4B Q4_K_M 5 GB Runs ~187 tok/s
Gemma 3n E4B 4B Q4_K_M 4.5 GB Runs ~208 tok/s
Qwen 3 4B 4B Q4_K_M 4.5 GB Runs ~208 tok/s
DeepSeek R1 7B 7B Q8_0 9 GB Runs ~104 tok/s
Falcon 3 7B 7B Q4_K_M 6.8 GB Runs ~138 tok/s
Mistral 7B 7B Q8_0 9 GB Runs ~104 tok/s
Qwen 2.5 7B 7B Q8_0 9 GB Runs ~104 tok/s
Qwen 2.5 Coder 7B 7B Q8_0 9 GB Runs ~104 tok/s
Qwen 2.5 VL 7B 7B Q4_K_M 7 GB Runs ~134 tok/s
Cogito 8B 8B Q4_K_M 7.5 GB Runs ~125 tok/s
DeepSeek R1 8B 8B Q4_K_M 7.5 GB Runs ~125 tok/s
Llama 3.1 8B 8B Q8_0 10 GB Runs ~94 tok/s
Nemotron 3 Nano 8B 8B Q4_K_M 7.5 GB Runs ~125 tok/s
Qwen 3 8B 8B Q4_K_M 7.5 GB Runs ~125 tok/s
Gemma 2 9B 9B Q8_0 11 GB Runs ~85 tok/s
Falcon 3 10B 10B Q4_K_M 8.5 GB Runs ~110 tok/s
Llama 3.2 Vision 11B 11B Q4_K_M 8.5 GB Runs ~110 tok/s
Gemma 3 12B 12B Q4_K_M 10.5 GB Runs ~89 tok/s
Mistral Nemo 12B 12B Q4_K_M 9.5 GB Runs ~99 tok/s
DeepSeek R1 14B 14B Q4_K_M 9.9 GB Runs ~95 tok/s
Phi-4 14B 14B Q4_K_M 9.9 GB Runs ~95 tok/s
Phi-4 Reasoning 14B 14B Q4_K_M 11 GB Runs ~85 tok/s
Qwen 2.5 14B 14B Q4_K_M 9.9 GB Runs ~95 tok/s
Qwen 2.5 Coder 14B 14B Q4_K_M 12 GB Runs ~78 tok/s
Qwen 3 14B 14B Q4_K_M 12 GB Runs ~78 tok/s
StarCoder2 15B 15B Q8_0 17 GB Runs ~55 tok/s
Codestral 22B 22B Q4_K_M 14.7 GB Runs ~64 tok/s
Devstral 24B 24B Q4_K_M 17 GB Runs ~55 tok/s
Magistral Small 24B 24B Q4_K_M 17 GB Runs ~55 tok/s
Mistral Small 3.1 24B 24B Q4_K_M 18 GB Runs ~52 tok/s
Gemma 2 27B 27B Q4_K_M 17.7 GB Runs ~53 tok/s
Gemma 3 27B 27B Q4_K_M 20 GB Runs ~47 tok/s
Qwen 3 30B-A3B (MoE) 30B Q4_K_M 22 GB Runs (tight) ~43 tok/s
Cogito 32B 32B Q4_K_M 21.5 GB Runs (tight) ~44 tok/s
DeepSeek R1 32B 32B Q4_K_M 20.7 GB Runs (tight) ~45 tok/s
Qwen 2.5 32B 32B Q4_K_M 20.7 GB Runs (tight) ~45 tok/s
QwQ 32B 32B Q4_K_M 21.5 GB Runs (tight) ~44 tok/s
Command R 35B 35B Q4_K_M 22.5 GB Runs (tight) ~42 tok/s
Qwen 2.5 Coder 32B 32B Q4_K_M 23 GB CPU Offload ~12 tok/s
Qwen 3 32B 32B Q4_K_M 23 GB CPU Offload ~12 tok/s
Mixtral 8x7B 47B Q4_K_M 29.7 GB CPU Offload ~10 tok/s
19 model(s) are too large for this hardware.