RTX 4090 (24 GB) vs RTX 3090 (24 GB)

The RTX 4090 (24 GB) runs 80 of 110 open models at 4-bit or better, the RTX 3090 (24 GB) runs 80. Usable memory: about 23 GB against 23 GB. The RTX 4090 (24 GB) has more memory bandwidth (1008 vs 936 GB/s), so models that fit both answer faster on it.

Data as of 2 November 2026. Speeds are estimates unless marked measured.

NVIDIA

GeForce RTX 4090

24 GB · 1 card

All models for this machine

NVIDIA

GeForce RTX 3090

24 GB · 1 card

All models for this machine

Running gpt-oss-20b

Runs

~169 tok/s

MXFP4 · 15.1 of 23 GB · ModelFit est.

Runs

~148 tok/s

MXFP4 · 15.1 of 23 GB · ModelFit est.

Single-user decode speed, estimated. Fastest machines for this model

Memory

24 GB

memory

24 GB

memory

~23 GB

usable for a model

~23 GB

usable for a model

1008 GB/s

memory bandwidth

936 GB/s

memory bandwidth

What it runs

80

of 110 models run at 4-bit or better

80

of 110 models run at 4-bit or better

Our pick

Gemma 4 31B

Q4_K_M · ~41 tok/s ModelFit est.

Gemma 4 31B

Q4_K_M · ~37 tok/s ModelFit est.

Largest current model that fits

Gemma 4 31B

30.7B at Q4_K_M

Gemma 4 31B

30.7B at Q4_K_M

Popular models

gpt-oss-120b

Too big

needs ~68 GB

Too big

needs ~68 GB

gpt-oss-20b

Runs

MXFP4 · ~169 tok/s

Runs

MXFP4 · ~148 tok/s

Qwen3.6-35B-A3B

Too big

needs ~24 GB

Too big

needs ~24 GB

Qwen3.8-27B

Runs

Q4_K_M · ~51 tok/s

Runs

Q4_K_M · ~45 tok/s

Gemma 4 26B-A4B

Runs

Q4_K_M · ~168 tok/s

Runs

Q4_K_M · ~147 tok/s

Llama 3.3 70B

Too big

needs ~44 GB

Too big

needs ~44 GB

Nemotron 3 Super 120B-A12B

Too big

needs ~67 GB

Too big

needs ~67 GB

DeepSeek-V4-Flash

Too big

needs ~177 GB

Too big

needs ~177 GB

Rent one by the hour

—

—

Use Change under a machine to pick another type, producer, model or variant. How the numbers are worked out