readygpu

NVIDIA GeForce RTX 4070 vs NVIDIA GeForce RTX 5090

For running models locally, priced in Germany. At 32k context the first runs 28 models fully in memory, the second 61.

In short

The NVIDIA GeForce RTX 5090 has more memory for models: 32 GB against 12 GB. At 32k context the NVIDIA GeForce RTX 5090 runs 61 of the models we track fully in memory and the NVIDIA GeForce RTX 4070 28; only the NVIDIA GeForce RTX 5090 holds Qwen 3.8 27B. On the models both run fully in memory, the NVIDIA GeForce RTX 5090 is about 3.6× as fast as the NVIDIA GeForce RTX 4070 (estimated, median across models).

Updated . Speeds are estimates; how they are worked out.

NVIDIA GeForce RTX 4070

Not priced here

Add to build
NVIDIA GeForce RTX 5090

6.199 €

eBay

Add to build

Specifications

SpecNVIDIA GeForce RTX 4070NVIDIA GeForce RTX 5090
GPUGeForce RTX 4070GeForce RTX 5090
Memory12 GB GDDR6X32 GB GDDR7
Memory bus192-bit512-bit
Memory bandwidth504.2 GB/s1,792 GB/s
FP16 compute29.146 TFLOPS209 TFLOPS
SoftwareCUDACUDA
Board power200 W575 W
PCIe generationGen 4Gen 5
PCIe lanesx16x16

What each runs

Estimated decode speed with the most precise quant that fits each one est.

ModelNVIDIA GeForce RTX 4070NVIDIA GeForce RTX 5090
Qwen 3.8 27Boffload35tok/sQ8_0
Qwen 3.6 27Boffload36tok/sQ8_0
Gemma 4 31Boffload39tok/sQ6_K
Qwen 3.6 35B-A3Boffload201tok/sQ6_K
Qwen 3.5 27Boffload36tok/sQ8_0
Gemma 4 26B-A4Boffload132tok/sQ8_0
Qwen 3.5 35B-A3Boffload212tok/sQ6_K
GLM 4.7 Flashoffload114tok/sQ6_K
Devstral Small 2 24Boffload37tok/sQ8_0
Nemotron 3.5 Lightning 30B-A3Boffload214tok/sUD-Q5_K_M
gpt-oss 20Boffload206tok/sMXFP4
Kimi Linear 48B-A3Boffload311tok/sQ4_K_M
Qwen3 Coder 30B-A3Boffload117tok/sQ6_K
Qwen3 32Boffload35tok/sQ5_K_M
Gemma 4 12B29tok/sQ6_K82tok/sQ8_0
Qwen 3.5 9B29tok/sQ8_0104tok/sQ8_0
Gemma 3 27Boffload35tok/sQ8_0
Mistral Small 3.2 24Boffload37tok/sQ8_0
DeepSeek R1 Distill Qwen 32Boffload35tok/sQ5_K_M
Ministral 3 14Boffload56tok/sQ8_0
Qwen3 14Boffload53tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B29tok/sQ5_K_M82tok/sQ8_0
Qwen3 8B29tok/sQ5_K_M82tok/sQ8_0
Phi-4 14Boffload50tok/sQ8_0
Qwen 3.5 4B56tok/sQ8_0198tok/sQ8_0
Ministral 3 8B29tok/sQ5_K_M82tok/sQ8_0
Granite 4.2 8B29tok/sQ4_K_M76tok/sQ8_0
Gemma 4 E4B34tok/sQ8_0121tok/sQ8_0
LFM2.5 8B-A1B70tok/sQ8_0249tok/sQ8_0
Llama 3.1 8B29tok/sQ6_K102tok/sQ6_K
Qwen2.5 0.5B335tok/sQ8_01190tok/sQ8_0
Qwen3.5 0.8B254tok/sQ8_0901tok/sQ8_0
Qwen3.5 2B168tok/sQ5_K_M598tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B110tok/sQ8_0392tok/sQ8_0
MiniCPM5 2B77tok/sQ8_0272tok/sQ8_0
Mellum2.1 12B A2.5B Thinking73tok/sQ5_K_M210tok/sQ8_0
Spark X2.5 4B73tok/smlx-4bit258tok/smlx-4bit
LFM2.5 2.6B64tok/sQ8_0227tok/sQ8_0
LFM2.5 VL 3B64tok/sQ8_0227tok/sQ8_0
FrogNano 4B 260954tok/sQ8_0193tok/sQ8_0
granite vision 4.1 4b50tok/sQ8_0176tok/sQ8_0
granite 4.2 3b48tok/sQ8_0169tok/sQ8_0
Ministral 3 3B Instruct 251244tok/sQ8_0156tok/sQ8_0
GLM 4.6V Flash42tok/sQ4_K_M98tok/sQ8_0
GLM 4 9B 041433tok/smlx-5bit116tok/smlx-5bit
DeepSeek R1 Distill Qwen 7B31tok/sQ8_0111tok/sQ8_0
Ministral 3 8B Reasoning 251229tok/sQ5_K_M82tok/sQ8_0
DeepSeek R1 Distill Llama 8B29tok/sQ6_K87tok/sQ8_0
Kimi VL A3B Thinking 2506offload162tok/sQ8_0
Kimi VL A3Boffload156tok/smlx-8bit
LFM2.5 24B A2Boffload139tok/sQ8_0
GLM 4.1V 9B Thinkingoffload85tok/smlx-8bit
Ministral 3 14B Reasoning 2512offload56tok/sQ8_0
DeepSeek R1 Distill Qwen 14Boffload50tok/sQ8_0
Muse Glimmer 30Boffload46tok/sQ6_K
Devstral Small 2507offload37tok/sQ8_0
Magistral Small 2506offload37tok/sQ8_0
Magistral Small 2507offload37tok/sQ8_0
Magistral Small 2509offload37tok/sQ8_0
Qwen3.8 27B OBLITERATEDoffload48tok/smlx-4bit
granite 4.2 30boffload38tok/sQ5_K_M
Qwen 3.5 122B-A10Bwon't runoffload
Nemotron 3 Super 120B-A12Bwon't runoffload
Qwen3 Coder Nextwon't runoffload
Mistral Small 4 119Bwon't runoffload
GLM 4.5 Airwon't runoffload
Qwen3 Next 80B-A3Bwon't runoffload
Llama 3.3 70Bwon't runoffload
DeepSeek R1 Distill Llama 70Bwon't runoffload
DeepSeek V4 Flashwon't runwon't run
GLM 5.3 Flashwon't runwon't run
Qwen 3.8 Flash Nextwon't runwon't run
MiniMax M2.7won't runwon't run
gpt-oss 120Bwon't runwon't run
DeepSeek R1won't runwon't run
DeepSeek R1 0528won't runwon't run
DeepSeek R1 Zerowon't runwon't run
DeepSeek V3won't runwon't run
DeepSeek V3 0324won't runwon't run
DeepSeek V3.1won't runwon't run
DeepSeek V3.1 Terminuswon't runwon't run
DeepSeek V3.2won't runwon't run
DeepSeek V4 Flash 0731won't runwon't run
DeepSeek V4 Flash Vision Expwon't runwon't run
Devstral 2 123B Instruct 2512won't runwon't run
GLM 4.5won't runwon't run
GLM 4.5Vwon't runwon't run
GLM 4.6won't runwon't run
GLM 4.6Vwon't runwon't run
GLM 4.7won't runwon't run
GLM 5won't runwon't run
GLM 5.1won't runwon't run
GLM 5.2won't runwon't run
GLM 5.3won't runwon't run
Kimi Dev 72Bwon't runwon't run
Kimi K2won't runwon't run
Kimi K2 Instruct 0905won't runwon't run
Kimi K2 Thinkingwon't runwon't run
Kimi K2.5won't runwon't run
MiniMax M2won't runwon't run
MiniMax M2.1won't runwon't run
MiniMax M2.5won't runwon't run
MiniMax M3won't runwon't run
Mistral Medium 3.5 128Bwon't runwon't run

Questions

Which is better for local AI, the NVIDIA GeForce RTX 4070 or the NVIDIA GeForce RTX 5090?
Choose the NVIDIA GeForce RTX 5090 to run bigger models. Choose the NVIDIA GeForce RTX 5090 for speed.
Which can run bigger models, the NVIDIA GeForce RTX 4070 or the NVIDIA GeForce RTX 5090?
At 32k context the NVIDIA GeForce RTX 5090 runs 61 of the models we track fully in memory and the NVIDIA GeForce RTX 4070 28; only the NVIDIA GeForce RTX 5090 holds Qwen 3.8 27B.
Which is faster for local AI, the NVIDIA GeForce RTX 4070 or the NVIDIA GeForce RTX 5090?
On the models both run fully in memory, the NVIDIA GeForce RTX 5090 is about 3.6× as fast as the NVIDIA GeForce RTX 4070 (estimated, median across models).

More comparisons

More hardware for local AI