Which is better for local AI, the NVIDIA GeForce RTX 4090 or the NVIDIA RTX PRO 4000 Blackwell?

Choose the NVIDIA GeForce RTX 4090 for speed.

NVIDIA GeForce RTX 4090 vs NVIDIA RTX PRO 4000 Blackwell for local AI (US)

In short

Both have 24 GB of memory for models; the NVIDIA GeForce RTX 4090 has 1,008 GB/s of bandwidth and the NVIDIA RTX PRO 4000 Blackwell 672 GB/s. At 128k context both run the same 31 models we track fully in memory. On the models both run fully in memory, the NVIDIA GeForce RTX 4090 is about 1.5× as fast as the NVIDIA RTX PRO 4000 Blackwell (estimated, median across models).

Updated October 26, 2026. Speeds are estimates; how they are worked out.

Specifications

Spec	NVIDIA GeForce RTX 4090	NVIDIA RTX PRO 4000 Blackwell
GPU	GeForce RTX 4090	RTX PRO 4000 Blackwell
Memory	24 GB GDDR6X	24 GB GDDR7
Memory bus	384-bit	192-bit
Memory bandwidth	1,008 GB/s	672 GB/s
FP16 compute	165 TFLOPS	46.9 TFLOPS
Software	CUDA	CUDA
Board power	450 W	145 W
PCIe generation	Gen 4	Gen 5
PCIe lanes	x16	x16

What each runs

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

Model	NVIDIA GeForce RTX 4090	NVIDIA RTX PRO 4000 Blackwell
Gemma 4 26B-A4B	64tok/sQ5_K_M	43tok/sQ5_K_M
gpt-oss 20B	66tok/sMXFP4	44tok/sMXFP4
Gemma 4 12B	41tok/sQ8_0	28tok/sQ8_0
Qwen 3.5 9B	45tok/sQ8_0	30tok/sQ8_0
Qwen 3.5 4B	71tok/sQ8_0	47tok/sQ8_0
Ministral 3 8B	27tok/sQ4_K_M	18tok/sQ4_K_M
Gemma 4 E4B	52tok/sQ8_0	35tok/sQ8_0
LFM2.5 8B-A1B	50tok/sQ8_0	33tok/sQ8_0
Llama 3.1 8B	27tok/sQ5_K_M	18tok/sQ5_K_M
Qwen2.5 0.5B	292tok/sQ8_0	195tok/sQ8_0
Qwen3.5 0.8B	256tok/sQ8_0	171tok/sQ8_0
Qwen3.5 2B	204tok/sQ5_K_M	136tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B	111tok/sQ8_0	74tok/sQ8_0
Mellum2.1 12B A2.5B Thinking	82tok/sQ8_0	55tok/sQ8_0
Spark X2.5 4B	79tok/smlx-4bit	53tok/smlx-4bit
MiniCPM5 2B	75tok/sQ8_0	50tok/sQ8_0
FrogNano 4B 2609	70tok/sQ8_0	47tok/sQ8_0
LFM2.5 2.6B	57tok/sQ8_0	38tok/sQ8_0
LFM2.5 VL 3B	57tok/sQ8_0	38tok/sQ8_0
Kimi VL A3B Thinking 2506	52tok/sQ8_0	35tok/sQ8_0
Kimi VL A3B	51tok/smlx-8bit	34tok/smlx-8bit
GLM 4 9B 0414	46tok/smlx-5bit	31tok/smlx-5bit
granite vision 4.1 4b	44tok/sQ8_0	29tok/sQ8_0
granite 4.2 3b	43tok/sQ8_0	29tok/sQ8_0
GLM 4.6V Flash	41tok/sQ8_0	27tok/sQ8_0
DeepSeek R1 Distill Qwen 7B	40tok/sQ8_0	27tok/sQ8_0
GLM 4.1V 9B Thinking	37tok/smlx-8bit	24tok/smlx-8bit
Ministral 3 3B Instruct 2512	36tok/sQ8_0	24tok/sQ8_0
Muse Glimmer 30B	29tok/sQ5_K_M	19tok/sQ5_K_M
DeepSeek R1 Distill Llama 8B	27tok/sQ5_K_M	18tok/sQ5_K_M
Ministral 3 8B Reasoning 2512	27tok/sQ4_K_M	18tok/sQ4_K_M
Qwen 3.8 27B	offload	offload
Qwen 3.5 122B-A10B	offload	offload
Qwen 3.6 27B	offload	offload
Gemma 4 31B	offload	offload
Mistral Small 4 119B	offload	offload
Qwen 3.6 35B-A3B	offload	offload
Qwen 3.5 27B	offload	offload
Qwen 3.5 35B-A3B	offload	offload
GLM 4.7 Flash	offload	offload
Devstral Small 2 24B	offload	offload
Nemotron 3.5 Lightning 30B-A3B	offload	offload
Kimi Linear 48B-A3B	offload	offload
Qwen3 Coder 30B-A3B	offload	offload
Gemma 3 27B	offload	offload
Mistral Small 3.2 24B	offload	offload
Ministral 3 14B	offload	offload
Qwen3 14B	offload	offload
DeepSeek R1 0528 Qwen3 8B	offload	offload
Qwen3 8B	offload	offload
Phi-4 14B	offload	offload
Granite 4.2 8B	offload	offload
LFM2.5 24B A2B	offload	offload
Qwen3.8 27B OBLITERATED	offload	offload
Ministral 3 14B Reasoning 2512	offload	offload
DeepSeek R1 Distill Qwen 14B	offload	offload
Devstral Small 2507	offload	offload
Magistral Small 2506	offload	offload
Magistral Small 2507	offload	offload
Magistral Small 2509	offload	offload
DeepSeek V4 Flash	won't run	won't run
GLM 5.3 Flash	won't run	won't run
Qwen 3.8 Flash Next	won't run	won't run
MiniMax M2.7	won't run	won't run
Nemotron 3 Super 120B-A12B	won't run	won't run
Qwen3 Coder Next	won't run	won't run
gpt-oss 120B	won't run	won't run
GLM 4.5 Air	won't run	won't run
Qwen3 Next 80B-A3B	won't run	won't run
Qwen3 32B	won't run	won't run
DeepSeek R1 Distill Qwen 32B	won't run	won't run
Llama 3.3 70B	won't run	won't run
DeepSeek R1	won't run	won't run
DeepSeek R1 0528	won't run	won't run
DeepSeek R1 Distill Llama 70B	won't run	won't run
DeepSeek R1 Zero	won't run	won't run
DeepSeek V3	won't run	won't run
DeepSeek V3 0324	won't run	won't run
DeepSeek V3.1	won't run	won't run
DeepSeek V3.1 Terminus	won't run	won't run
DeepSeek V3.2	won't run	won't run
DeepSeek V4 Flash 0731	won't run	won't run
DeepSeek V4 Flash Vision Exp	won't run	won't run
Devstral 2 123B Instruct 2512	won't run	won't run
GLM 4.5	won't run	won't run
GLM 4.5V	won't run	won't run
GLM 4.6	won't run	won't run
GLM 4.6V	won't run	won't run
GLM 4.7	won't run	won't run
GLM 5	won't run	won't run
GLM 5.1	won't run	won't run
GLM 5.2	won't run	won't run
GLM 5.3	won't run	won't run
Kimi Dev 72B	won't run	won't run
Kimi K2	won't run	won't run
Kimi K2 Instruct 0905	won't run	won't run
Kimi K2 Thinking	won't run	won't run
Kimi K2.5	won't run	won't run
MiniMax M2	won't run	won't run
MiniMax M2.1	won't run	won't run
MiniMax M2.5	won't run	won't run
MiniMax M3	won't run	won't run
Mistral Medium 3.5 128B	won't run	won't run
granite 4.2 30b	won't run	won't run

Questions

Which is better for local AI, the NVIDIA GeForce RTX 4090 or the NVIDIA RTX PRO 4000 Blackwell?: Choose the NVIDIA GeForce RTX 4090 for speed.
Which can run bigger models, the NVIDIA GeForce RTX 4090 or the NVIDIA RTX PRO 4000 Blackwell?: At 128k context both run the same 31 models we track fully in memory.
Which is faster for local AI, the NVIDIA GeForce RTX 4090 or the NVIDIA RTX PRO 4000 Blackwell?: On the models both run fully in memory, the NVIDIA GeForce RTX 4090 is about 1.5× as fast as the NVIDIA RTX PRO 4000 Blackwell (estimated, median across models).

NVIDIA GeForce RTX 4090 vs NVIDIA RTX PRO 4000 Blackwell

In short

Specifications

What each runs

Questions

More comparisons

More hardware for local AI