readygpu

NVIDIA GeForce RTX 3090 vs NVIDIA RTX PRO 4500 Blackwell

For running models locally, priced in the UK. At 64k context the first runs 48 models fully in memory, the second 58.

In short

The NVIDIA RTX PRO 4500 Blackwell has more memory for models: 32 GB against 24 GB. At 64k context the NVIDIA RTX PRO 4500 Blackwell runs 58 of the models we track fully in memory and the NVIDIA GeForce RTX 3090 48; only the NVIDIA RTX PRO 4500 Blackwell holds Gemma 4 31B. On the models both run fully in memory, the NVIDIA GeForce RTX 3090 and the NVIDIA RTX PRO 4500 Blackwell are about as fast as each other (estimated).

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

NVIDIA GeForce RTX 3090

£1,330

Amazon

Add to build
NVIDIA RTX PRO 4500 Blackwell

Not priced here

Add to build

Specifications

SpecNVIDIA GeForce RTX 3090NVIDIA RTX PRO 4500 Blackwell
GPUGeForce RTX 3090RTX PRO 4500 Blackwell
Memory24 GB GDDR6X32 GB GDDR7
Memory bus384-bit256-bit
Memory bandwidth936 GB/s896 GB/s
FP16 compute71 TFLOPS54.94 TFLOPS
SoftwareCUDACUDA
Board power350 W200 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 4500 Blackwell
Qwen 3.8 27B27tok/sQ4_K_M20tok/sQ6_K
Qwen 3.6 27B24tok/sQ5_K_M21tok/sQ6_K
Gemma 4 31Boffload18tok/sQ6_K
Qwen 3.6 35B-A3B107tok/sQ4_K_M84tok/sQ6_K
Qwen 3.5 27B27tok/sQ4_K_M21tok/sQ6_K
Gemma 4 26B-A4B72tok/sUD-Q5_K_M58tok/sQ8_0
Qwen 3.5 35B-A3B107tok/sQ4_K_M87tok/sQ6_K
GLM 4.7 Flash53tok/sQ4_K_M44tok/sQ6_K
Devstral Small 2 24Boffload19tok/sQ6_K
Nemotron 3.5 Lightning 30B-A3B131tok/sMXFP4100tok/sUD-Q5_K_M
gpt-oss 20B86tok/sMXFP483tok/sMXFP4
Kimi Linear 48B-A3Boffload139tok/sQ4_K_M
Qwen3 Coder 30B-A3Boffload37tok/sQ6_K
Gemma 4 12B41tok/sQ8_040tok/sQ8_0
Qwen 3.5 9B50tok/sQ8_047tok/sQ8_0
Gemma 3 27B26tok/sQ4_K_M20tok/sQ6_K
Mistral Small 3.2 24Boffload19tok/sQ6_K
Ministral 3 14B27tok/sQ6_K22tok/sQ8_0
Qwen3 14B25tok/sQ6_K21tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B32tok/sQ8_030tok/sQ8_0
Qwen3 8B32tok/sQ8_030tok/sQ8_0
Phi-4 14B24tok/sQ5_K_M19tok/sQ8_0
Qwen 3.5 4B87tok/sQ8_083tok/sQ8_0
Ministral 3 8B32tok/sQ8_031tok/sQ8_0
Granite 4.2 8B29tok/sQ8_028tok/sQ8_0
Gemma 4 E4B57tok/sQ8_055tok/sQ8_0
LFM2.5 8B-A1B81tok/sQ8_078tok/sQ8_0
Llama 3.1 8B38tok/sQ6_K37tok/sQ6_K
Qwen2.5 0.5B434tok/sQ8_0416tok/sQ8_0
Qwen3.5 0.8B355tok/sQ8_0340tok/sQ8_0
Qwen3.5 2B257tok/sQ5_K_M246tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B154tok/sQ8_0147tok/sQ8_0
MiniCPM5 2B106tok/sQ8_0101tok/sQ8_0
Spark X2.5 4B105tok/smlx-4bit101tok/smlx-4bit
Mellum2.1 12B A2.5B Thinking96tok/sQ8_092tok/sQ8_0
FrogNano 4B 260985tok/sQ8_082tok/sQ8_0
LFM2.5 2.6B84tok/sQ8_081tok/sQ8_0
LFM2.5 VL 3B84tok/sQ8_081tok/sQ8_0
Kimi VL A3B Thinking 250668tok/sQ8_065tok/sQ8_0
Kimi VL A3B66tok/smlx-8bit63tok/smlx-8bit
granite vision 4.1 4b65tok/sQ8_062tok/sQ8_0
granite 4.2 3b63tok/sQ8_060tok/sQ8_0
Ministral 3 3B Instruct 251255tok/sQ8_052tok/sQ8_0
GLM 4 9B 041453tok/smlx-5bit51tok/smlx-5bit
LFM2.5 24B A2B50tok/sQ5_K_M45tok/sQ8_0
DeepSeek R1 Distill Qwen 7B49tok/sQ8_047tok/sQ8_0
GLM 4.6V Flash46tok/sQ8_044tok/sQ8_0
GLM 4.1V 9B Thinking40tok/smlx-8bit38tok/smlx-8bit
DeepSeek R1 Distill Llama 8B34tok/sQ8_032tok/sQ8_0
Ministral 3 8B Reasoning 251232tok/sQ8_031tok/sQ8_0
Muse Glimmer 30B28tok/sQ5_K_M23tok/sQ6_K
Ministral 3 14B Reasoning 251227tok/sQ6_K22tok/sQ8_0
DeepSeek R1 Distill Qwen 14B25tok/sQ5_K_M19tok/sQ8_0
Devstral Small 2507offload20tok/sQ5_K_M
Magistral Small 2506offload19tok/sQ6_K
Magistral Small 2507offload20tok/sQ5_K_M
Magistral Small 2509offload20tok/sQ5_K_M
Qwen3.8 27B OBLITERATEDoffload22tok/smlx-4bit
Qwen 3.5 122B-A10Boffloadoffload
Qwen3 Coder Nextoffloadoffload
Nemotron 3 Super 120B-A12Bwon't runoffload
Mistral Small 4 119Boffloadoffload
GLM 4.5 Airwon't runoffload
Qwen3 Next 80B-A3Boffloadoffload
Qwen3 32Boffloadoffload
DeepSeek R1 Distill Qwen 32Boffloadoffload
granite 4.2 30boffloadoffload
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
Llama 3.3 70Bwon't runwon't run
DeepSeek R1won't runwon't run
DeepSeek R1 0528won't runwon't run
DeepSeek R1 Distill Llama 70Bwon'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 3090 or the NVIDIA RTX PRO 4500 Blackwell?
Choose the NVIDIA RTX PRO 4500 Blackwell to run bigger models.
Which can run bigger models, the NVIDIA GeForce RTX 3090 or the NVIDIA RTX PRO 4500 Blackwell?
At 64k context the NVIDIA RTX PRO 4500 Blackwell runs 58 of the models we track fully in memory and the NVIDIA GeForce RTX 3090 48; only the NVIDIA RTX PRO 4500 Blackwell holds Gemma 4 31B.
Which is faster for local AI, the NVIDIA GeForce RTX 3090 or the NVIDIA RTX PRO 4500 Blackwell?
On the models both run fully in memory, the NVIDIA GeForce RTX 3090 and the NVIDIA RTX PRO 4500 Blackwell are about as fast as each other (estimated).

More comparisons

More hardware for local AI