Datasheet for local AI49 models98 GPUsData read 2026-09-25
No adsNo tracking
Check my GPU →

RTX 5090 32 GB for local AI

Which image and video models run on the RTX 5090 32 GB, which file to download for each, and how much VRAM they need.

NVIDIA32 GB GDDR7 · 512-bit · 1792 GB/s · BlackwellData 2026-09-25
VRAM32 GB
MemoryGDDR7
Bus512-bit
Bandwidth1792 GB/s
ArchitectureBlackwell
Launched2025-01
Launch price$1,999
Runs well47 of 49

Specs: www.nvidia.com · launch: nvidianews.nvidia.com

01What runs on it

ModelVerdictBest fileSizeNeeded
image models
FLUX.1 [dev]Runs well16-bit23.8 GB26.1 GB
FLUX.1 [schnell]Runs well16-bit23.8 GB26.1 GB
FLUX.1 Kontext [dev]Runs well16-bit23.8 GB26.4 GB
FLUX.1 Krea [dev]Runs well16-bit23.8 GB26.1 GB
FLUX.1 Fill [dev]Runs well16-bit23.8 GB26.4 GB
FLUX.2 [dev]RunsQ6_K27.4 GB30.7 GB
FLUX.2 [klein] 9BRuns well16-bit18.2 GB20.5 GB
FLUX.2 [klein] 4BRuns well16-bit7.8 GB9.6 GB
Krea 2 (Turbo)Runs well16-bit26.3 GB28.9 GB
Qwen-ImageRuns wellFP820.4 GB23.2 GB
Qwen-Image-Edit (2511)Runs wellFP820.5 GB23.5 GB
Qwen-Image 2.1Runs well16-bit14.2 GB16.5 GB
Z-Image TurboRuns well16-bit12.3 GB14.3 GB
Z-Image (base)Runs well16-bit12.3 GB14.3 GB
Ideogram 4Runs wellFP89.3 GB20.9 GB
Boogu-Image (Turbo)Runs well16-bit20.6 GB23.2 GB
ERNIE-Image (Turbo)Runs well16-bit16.1 GB18.4 GB
HiDream-O1-ImageRuns well16-bit16.4 GB19.2 GB
Mage-Flow (Microsoft)Runs well16-bit8.2 GB10.0 GB
Lumina Image 2.0Runs well16-bit5.2 GB7.0 GB
HiDream-I1 (Full)Runs wellFP817.1 GB19.9 GB
HiDream-I1 (Dev)Runs wellFP817.1 GB19.9 GB
Stable Diffusion 3.5 LargeRuns well16-bit16.5 GB18.8 GB
Stable Diffusion 3.5 MediumRuns well16-bit5.1 GB6.9 GB
Chroma1-HDRuns well16-bit17.8 GB20.1 GB
SDXL 1.0Runs well16-bit6.9 GB7.1 GB
Illustrious XL / Pony (SDXL anime)Runs well16-bit6.9 GB7.1 GB
Stable Diffusion 1.5Runs well16-bit2.1 GB3.3 GB
HunyuanImage 2.1Runs wellFP817.4 GB20.7 GB
video models
Wan 2.1 T2V 14BRuns wellFP814.3 GB18.6 GB
Wan 2.1 T2V 1.3BRuns well16-bit2.8 GB5.6 GB
Wan 2.1 I2V 14B 480PRuns wellFP816.4 GB20.7 GB
Wan 2.1 I2V 14B 720PRuns wellFP816.4 GB23.2 GB
Wan 2.1 VACE 14BRuns wellQ8_018.7 GB23.5 GB
Wan 2.2 T2V A14BRuns wellFP814.3 GB18.6 GB
Wan 2.2 I2V A14BRuns wellFP814.3 GB18.6 GB
Wan 2.2 TI2V 5BRuns well16-bit10.0 GB13.8 GB
Wan 2.2 Animate 14BRuns wellFP817.3 GB22.6 GB
Wan Animate 2 (14B)Runs wellQ8_018.1 GB23.4 GB
Wan 2.2 S2V 14BRuns wellFP816.4 GB21.2 GB
SCAIL-2 (character animation)Runs wellFP817.7 GB23.0 GB
HunyuanVideo (13B, original)Runs well16-bit25.6 GB29.9 GB
LTX-Video 13B (0.9.8)Runs wellFP815.7 GB20.0 GB
HunyuanVideo 1.5Runs well16-bit16.7 GB21.0 GB
LTX-2 (19B)Runs wellQ8_020.4 GB25.2 GB
LTX-2.3 (22B)Runs wellQ8_022.8 GB27.6 GB
LTX-2.5 (22B)Runs wellQ8_023.6 GB28.4 GB
MiniMax H3 (33B)RunsQ5_K_M23.9 GB29.7 GB
MiniMax H3 PrunedRuns wellFP821.0 GB26.8 GB

Calculated from real file sizes plus working memory. How the numbers work.

02Good to know

The RTX 5090 32 GB is a Blackwell GPU with FP8 and FP4 hardware: ComfyUI computes Comfy-Org's FP8 files natively here, and NVFP4 files (where a model offers them) are faster still.

03Measured and reported results

LabelModelSetupResultPeak VRAMDateSource
reportedFLUX.1 devfp8 (ComfyUI template)
“Getting 8.78s at 2.38it/s for 3 runs.”
ComfyUI 'GPU Benchmark Flux DEV fp8' thread: stock Flux dev fp8 workflow template, time of 2nd/3rd run (no loading); resolution/steps not stated in thread; Inno3D RTX 5090 X3 OC
8.78 s / image · 2.38 it/s—2025-08-05github.com →
reportedHunyuanVideo 1.5720p model · 848x480
“I was able to generate 5 seconds video in 284s at 24fps using the 720p model at 848*480”
ComfyUI; 5 s video at 24 fps; steps not stated
284 s / clip—2025-11-22huggingface.co →
reportedIllustrious / PonyIllustrious-XL-v2.0 · 1024x1024
“NVIDIA RTX 5090 | 8.95it/s | 0.11s/it | CUDA 12.8 | ComfyUI (Unknown) | Windows 11 24H2”
Community speed list; ComfyUI, Euler/Normal, CFG 8, square 1:1, batch speed only (no per-image time); date = list last-updated date; SDXL 1024px section
0.11 s/it—2025-11-29huggingface.co →
reportedLTX-2LTX-2 19B NVFP4 · 720p
“generating 143 frames at 720p takes about 66 seconds end‑to‑end in ComfyUI”
Issue says this is ~30-40% slower than NVIDIA's expected 40-45 s; steps not stated
66 s / clip (143 frames)—2026-01-08github.com →
reportedMiniMax H3 Prunedminimax_h3_fl2va_pruned_nvfp4 · 864x480 · 10 steps
“175 s for a 864×480 ten-second clip, 26,914 MiB peak VRAM.”
ComfyUI 0.30.1, 500 W power cap; 26,914 MiB = 26.28 GiB; INT8-ConvRot = 185 s / 28,581 MiB
175 s / clip (243 frames)26.28 GB2026-08-04ai-muninn.com →
reportedZ-Image Turbofp8_e4m3fn (per log)
“without SA: 0.95 seg”
'seg' = segundos (seconds) per image; SageAttention variants gave 0.87-0.91 s; resolution/steps not stated
0.95 s / image—2025-11-27github.com →

Reported results are other people's numbers, copied as published, with a link. Settings, drivers and ComfyUI versions differ, so compare them with care. Send yours.