From the RTX 3090 24 GB to the RTX 5090 32 GB: what changes for local AI
The change you would notice first: LTX-2.5 (22B) goes from a Q6_K file to Q8_0, entirely in VRAM. 3 other models reach 8-bit or better too, and 9 more get a better file.
4newly run well
9better file
37no change
+8GB more VRAM
01Newly runs well
| Model | On the RTX 3090 24 GB | On the RTX 5090 32 GB | Needed |
|---|---|---|---|
| LTX-2 (19B) | Runs Q6_K | Runs well Q8_0 | 25.2 GB |
| LTX-2.3 (22B) | Runs Q6_K | Runs well Q8_0 | 27.6 GB |
| LTX-2.5 (22B) | Runs Q6_K | Runs well Q8_0 | 28.4 GB |
| MiniMax H3 Pruned | Runs Q6_K | Runs well FP8 | 26.8 GB |
“Runs well” means a 16-bit or 8-bit file fits entirely in VRAM.
02A better file, or less offloading
| Model | On the RTX 3090 24 GB | On the RTX 5090 32 GB | Needed |
|---|---|---|---|
| FLUX.1 [dev] | Runs well Q8_0 | Runs well 16-bit | 26.1 GB |
| FLUX.1 [schnell] | Runs well Q8_0 | Runs well 16-bit | 26.1 GB |
| FLUX.1 Kontext [dev] | Runs well Q8_0 | Runs well 16-bit | 26.4 GB |
| FLUX.1 Krea [dev] | Runs well Q8_0 | Runs well 16-bit | 26.1 GB |
| FLUX.1 Fill [dev] | Runs well Q8_0 | Runs well 16-bit | 26.4 GB |
| FLUX.2 [dev] | Runs Q4_K_M | Runs Q6_K | 30.7 GB |
| Krea 2 (Turbo) | Runs well Q8_0 | Runs well 16-bit | 28.9 GB |
| HunyuanVideo (13B, original) | Runs well Q8_0 | Runs well 16-bit | 29.9 GB |
| MiniMax H3 (33B) | Tight Q3_K_M | Runs Q5_K_M | 29.7 GB |
03No change in what fits
FLUX.2 klein 9B Runs wellFLUX.2 klein 4B Runs wellQwen-Image Runs wellQwen-Image-Edit Runs wellQwen-Image 2.1 Runs wellZ-Image Turbo Runs wellZ-Image Runs wellIdeogram 4 Runs wellBoogu-Image Runs wellERNIE-Image Runs wellHiDream-O1 Runs wellMage-Flow Runs wellMing-Image Runs wellLumina 2.0 Runs wellHiDream-I1 Full Runs wellHiDream-I1 Runs wellSD 3.5 Large Runs wellSD 3.5 Medium Runs wellChroma1-HD Runs wellSDXL Runs wellIllustrious / Pony Runs wellSD 1.5 Runs wellHunyuanImage 2.1 Runs wellWan 2.1 14B Runs wellWan 2.1 1.3B Runs wellWan 2.1 I2V 480P Runs wellWan 2.1 I2V 720P Runs wellWan VACE 14B Runs wellWan 2.2 T2V Runs wellWan 2.2 I2V Runs wellWan 2.2 5B Runs wellWan 2.2 Animate Runs wellWan Animate 2 Runs wellWan 2.2 S2V Runs wellSCAIL-2 Runs wellLTX-Video 13B Runs wellHunyuanVideo 1.5 Runs well
Same verdict and same file on both cards. Speed can still differ.
04Speed and features
Memory bandwidth goes from 936 to 1792 GB/s (×1.91). A rough guide to speed, not a benchmark. The RTX 5090 32 GB computes FP8 natively, so FP8 files run faster there than on the RTX 3090 24 GB, which only uses them to save memory. It also adds FP4 (NVFP4) hardware, for models that ship NVFP4 files.