readygpu

NVIDIA GeForce RTX 3090 vs NVIDIA RTX PRO 5000 Blackwell 72GB

For running models locally, priced in the UK. At 32k context the first runs 57 models fully in memory, the second 72.

In short

The NVIDIA RTX PRO 5000 Blackwell 72GB has more memory for models: 72 GB against 24 GB. At 32k context the NVIDIA RTX PRO 5000 Blackwell 72GB runs 72 of the models we track fully in memory and the NVIDIA GeForce RTX 3090 57; only the NVIDIA RTX PRO 5000 Blackwell 72GB holds Qwen 3.5 122B-A10B. On the models both run fully in memory, the NVIDIA RTX PRO 5000 Blackwell 72GB is about 1.4× as fast as the NVIDIA GeForce RTX 3090 (estimated, median across models).

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

NVIDIA GeForce RTX 3090

£1,330

Amazon

Add to build
NVIDIA RTX PRO 5000 Blackwell 72GB

Not priced here

Add to build

Specifications

SpecNVIDIA GeForce RTX 3090NVIDIA RTX PRO 5000 Blackwell 72GB
GPUGeForce RTX 3090RTX PRO 5000 Blackwell 72GB
Memory24 GB GDDR6X72 GB GDDR7
Memory bus384-bit384-bit
Memory bandwidth936 GB/s1,344 GB/s
FP16 compute71 TFLOPS73.69 TFLOPS
SoftwareCUDACUDA
Board power350 W300 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 3090NVIDIA RTX PRO 5000 Blackwell 72GB
Qwen 3.8 27B25tok/sQ5_K_M27tok/sQ8_0
Qwen 3.5 122B-A10Boffload92tok/sQ3_K_M
Qwen 3.6 27B27tok/sQ5_K_M27tok/sQ8_0
Gemma 4 31B27tok/sQ4_K_M23tok/sQ8_0
Qwen3 Coder Nextoffload123tok/sQ6_K
Nemotron 3 Super 120B-A12Bwon't run78tok/sUD-Q3_K_M
gpt-oss 120Bwon't run115tok/sMXFP4
Mistral Small 4 119Boffload98tok/sUD-Q4_K_M
Qwen 3.6 35B-A3B134tok/sQ4_K_M131tok/sQ8_0
Qwen 3.5 27B27tok/sQ5_K_M27tok/sQ8_0
Gemma 4 26B-A4B78tok/sUD-Q6_K99tok/sQ8_0
GLM 4.5 Airoffload37tok/sQ3_K_M
Qwen 3.5 35B-A3B135tok/sQ4_K_M131tok/sQ8_0
GLM 4.7 Flash65tok/sQ5_K_M71tok/sQ8_0
Devstral Small 2 24B26tok/sQ5_K_M27tok/sQ8_0
Qwen3 Next 80B-A3Boffload152tok/sQ6_K
Nemotron 3.5 Lightning 30B-A3B142tok/sMXFP4140tok/sUD-Q6_K_XL
gpt-oss 20B108tok/sMXFP4155tok/sMXFP4
Qwen3 Coder 30B-A3B69tok/sQ4_K_M78tok/sQ8_0
Kimi Linear 48B-A3Boffload143tok/sQ8_0
Qwen3 32Boffload19tok/sQ8_0
Gemma 4 12B43tok/sQ8_062tok/sQ8_0
Qwen 3.5 9B55tok/sQ8_078tok/sQ8_0
Mistral Small 3.2 24B26tok/sQ5_K_M27tok/sQ8_0
Gemma 3 27B26tok/sQ5_K_M26tok/sQ8_0
DeepSeek R1 Distill Qwen 32Boffload19tok/sQ8_0
Llama 3.3 70Bwon't run14tok/sQ5_K_M
Ministral 3 14B29tok/sQ8_042tok/sQ8_0
Qwen3 14B28tok/sQ8_040tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B43tok/sQ8_062tok/sQ8_0
Qwen3 8B43tok/sQ8_062tok/sQ8_0
Phi-4 14B26tok/sQ8_037tok/sQ8_0
Qwen 3.5 4B104tok/sQ8_0149tok/sQ8_0
Ministral 3 8B43tok/sQ8_061tok/sQ8_0
Granite 4.2 8B39tok/sQ8_057tok/sQ8_0
Gemma 4 E4B63tok/sQ8_091tok/sQ8_0
LFM2.5 8B-A1B130tok/sQ8_0187tok/sQ8_0
Llama 3.1 8B53tok/sQ6_K77tok/sQ6_K
Qwen2.5 0.5B622tok/sQ8_0892tok/sQ8_0
Qwen3.5 0.8B471tok/sQ8_0676tok/sQ8_0
Qwen3.5 2B313tok/sQ5_K_M449tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B205tok/sQ8_0294tok/sQ8_0
MiniCPM5 2B142tok/sQ8_0204tok/sQ8_0
Spark X2.5 4B135tok/smlx-4bit194tok/smlx-4bit
LFM2.5 2.6B119tok/sQ8_0171tok/sQ8_0
LFM2.5 VL 3B119tok/sQ8_0171tok/sQ8_0
Mellum2.1 12B A2.5B Thinking110tok/sQ8_0158tok/sQ8_0
FrogNano 4B 2609101tok/sQ8_0145tok/sQ8_0
granite vision 4.1 4b92tok/sQ8_0132tok/sQ8_0
granite 4.2 3b88tok/sQ8_0127tok/sQ8_0
Kimi VL A3B Thinking 250685tok/sQ8_0121tok/sQ8_0
Kimi VL A3B81tok/smlx-8bit117tok/smlx-8bit
Ministral 3 3B Instruct 251281tok/sQ8_0117tok/sQ8_0
LFM2.5 24B A2B78tok/sQ6_K104tok/sQ8_0
GLM 4 9B 041461tok/smlx-5bit87tok/smlx-5bit
DeepSeek R1 Distill Qwen 7B58tok/sQ8_084tok/sQ8_0
GLM 4.6V Flash51tok/sQ8_074tok/sQ8_0
DeepSeek R1 Distill Llama 8B45tok/sQ8_065tok/sQ8_0
GLM 4.1V 9B Thinking44tok/smlx-8bit64tok/smlx-8bit
Ministral 3 8B Reasoning 251243tok/sQ8_061tok/sQ8_0
Ministral 3 14B Reasoning 251229tok/sQ8_042tok/sQ8_0
DeepSeek R1 Distill Qwen 14B26tok/sQ8_038tok/sQ8_0
Devstral Small 250726tok/sQ5_K_M27tok/sQ8_0
Magistral Small 250626tok/sQ5_K_M27tok/sQ8_0
Magistral Small 250726tok/sQ5_K_M27tok/sQ8_0
Magistral Small 250926tok/sQ5_K_M27tok/sQ8_0
Qwen3.8 27B OBLITERATED25tok/smlx-4bit36tok/smlx-4bit
Muse Glimmer 30B24tok/sQ6_K25tok/sQ8_0
granite 4.2 30boffload21tok/sQ8_0
DeepSeek R1 Distill Llama 70Bwon't run14tok/sQ5_K_M
GLM 4.5Vwon't run34tok/sQ4_K_M
Kimi Dev 72Bwon't run11tok/sQ6_K
DeepSeek V4 Flashwon't runoffload
Qwen 3.8 Flash Nextwon't runoffload
MiniMax M2.7won't runoffload
Devstral 2 123B Instruct 2512won't runoffload
GLM 4.6Vwon't runoffload
Mistral Medium 3.5 128Bwon't runoffload
GLM 5.3 Flashwon'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
GLM 4.5won't runwon't run
GLM 4.6won'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 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

Questions

Which is better for local AI, the NVIDIA GeForce RTX 3090 or the NVIDIA RTX PRO 5000 Blackwell 72GB?
Choose the NVIDIA RTX PRO 5000 Blackwell 72GB to run bigger models. Choose the NVIDIA RTX PRO 5000 Blackwell 72GB for speed.
Which can run bigger models, the NVIDIA GeForce RTX 3090 or the NVIDIA RTX PRO 5000 Blackwell 72GB?
At 32k context the NVIDIA RTX PRO 5000 Blackwell 72GB runs 72 of the models we track fully in memory and the NVIDIA GeForce RTX 3090 57; only the NVIDIA RTX PRO 5000 Blackwell 72GB holds Qwen 3.5 122B-A10B.
Which is faster for local AI, the NVIDIA GeForce RTX 3090 or the NVIDIA RTX PRO 5000 Blackwell 72GB?
On the models both run fully in memory, the NVIDIA RTX PRO 5000 Blackwell 72GB is about 1.4× as fast as the NVIDIA GeForce RTX 3090 (estimated, median across models).

More comparisons

More hardware for local AI