Best video models to run locally on a 64GB graphics card
A leaderboard of the best video models you can run locally on a 64GB graphics card (with 32GB of system RAM). MiniMax H3 is #1 right now, followed by Wan 2.2 I2V A14B and LTX-2.5. Models are ranked by one score that combines 55 published benchmarks. Each model shows the download that fits and the command to run it.
46 models · scores combine 55 benchmarks · updated how it works · find a leaderboard · benchmark your machine
Models for 32GB RAM · 64GB graphics card → ~64GB fast · local only · reset
+ more
$ terminal
↗ share
detecting your OS…
Score vs memory to run
x
show
● open weights ◆ API only ─┘ best score for the memory · x: GB of RAM / VRAM for the smallest download (log)
| # | model | score↓ | where↓ | speed↓ | sources |
|---|---|---|---|---|---|
| 6 |  MiniMax H3 · gguf-q4_k_m fl2va + Qwen3-VL 32B q4_k_m (sd.cpp) | 78±1 | local |    +2 | |
| 42 |  Wan 2.2 I2V A14B · gguf-q8_0 ×2 experts + umt5 q8_0 (sd.cpp) | 56±2 | local | +2 | |
| 53 |  LTX-2.5 · gguf-q8_0 22B dev + Gemma 4 12B q8_0 (sd.cpp) | 53±6* | local | | |
| 55 |  HunyuanVideo 1.5 (I2V) · bf16 480p I2V + model CPU offload | 53±2 | local (slow) | +2 | |
| 57 |  Kandinsky 5.0 Pro · bf16 + model CPU offload | 52±2* | local (slow) | | |
| 59 | HunyuanVideo 1.5 · fp16 720p T2V (sd.cpp) | 52±2 | local | | |
| 60 |  Cosmos3-Nano · bf16 | 52±3 | local | | |
| 64 | Kandinsky 6.0 Video Lite · bf16 + model CPU offload | 51±7*† | local | | |
| 67 | Wan 2.2 A14B · gguf-q8_0 ×2 experts + umt5 q8_0 (sd.cpp) | 50±2 | local | | |
| 71 |  Open-Sora Plan v1.3 (T2V) · bf16 | 49±7* | local | | |
| 80 | Wan 2.1 14B · bf16 | 47±2 | local |  | |
| 81 | Wan 2.2 TI2V 5B · fp16 + umt5 q8_0 (sd.cpp) | 47±2 | local | +3 | |
| 82 | Kandinsky 5.0 T2V Lite · bf16 | 47±2* | local | | |
| 85 |  CogVideoX-5B-I2V · bf16 | 46±4 | local | | |
| 86 | Cosmos-1.0-Diffusion-7B-Text2World · bf16 | 46±3 | local | | |
| 90 | HunyuanVideo-I2V · bf16 720p + CPU offload (official script) | 45±4 | local | | |
| 91 | CogVideoX-5B · bf16 | 44±2 | local | | |
| 95 |  Open-Sora 2.0 · bf16 256px direct text-to-video | 42±5* | local | | |
| 96 |  Pyramid Flow (SD3) · bf16 | 42±3* | local | | |
| 99 | Wan 2.1 T2V 1.3B · fp16 + umt5 q8_0 (sd.cpp) | 40±4 | local | | |
| 100 |  Mochi 1 · bf16 | 40±2 | local | | |
| 102 | Open-Sora Plan v1.3 (I2V) · bf16 | 39±7* | local | | |
| 105 | HunyuanVideo · bf16 | 39±3 | local | | |
| 107 |  T2V-Turbo-v2 · fp32 | 38±5* | local | | |
| 109 | T2V-Turbo (VC2) · fp32 | 38±3† | local | | |
| 112 | LTX-2.3 · gguf-ud-q4_k_m 22B dev + Gemma 3 12B q4 (sd.cpp) | 36±3 | local | | |
| 113 | CogVideoX1.5-5B · bf16 + sequential CPU offload | 36±2 | local | | |
| 114 | CogVideoX-2B · fp16 | 36±2 | local | | |
| 116 |  Allegro · bf16 | 35±2 | local | | |
| 118 |  VideoCrafter2 · fp32 | 34±2 | local | | |
| 121 | Open-Sora 1.1 · bf16 | 32±7* | local | | |
| 122 |  LaVie (base T2V) · fp16 | 32±2 | local | | |
| 123 | LTX-Video 0.9 (2B) · openvino-int8 (OpenVINO GenAI, one-time export) | 32±3* | local | | |
| 125 |  NOVA-0.6B T2V · fp16 | 29±3* | local | | |
| 126 | ZeroScope v2 576w · fp16 | 27±3 | local | | |
| 127 |  Hotshot-XL · fp16 | 27±2 | local | | |
| 130 | Open-Sora 1.2 · bf16 | 27±3 | local | | |
| 133 | VideoCrafter1 T2V · official script (checkpoint as stored) | 26±2 | local | | |
| 134 |  Show-1 · fp16 + model CPU offload (official run_inference.py) | 25±2 | local (slow) | | |
| 135 |  LVDM (text-to-video) · fp32 | 25±3* | local | | |
| 136 |  ModelScope T2V 1.7B · fp16 | 25±2 | local | | |
| 138 | ZeroScope v2 XL (576w → XL pipeline) · fp16 + model CPU offload (576w base, then XL vid2vid) | 24±2 | local | | |
| 139 | CogVideo 9B (original) · fp16 (--fp16), both stages | 23±7* | local | | |
| — | AnimateDiff SD 1.5 v2 motion adapter · fp32 motion module + SD 1.5 (sd.cpp) | no reviewed results | local | ||
| — |  Stable Video Diffusion (SVD, 14 frames) · fp16 · CPU offload | no reviewed results | local | ||
| — | Stable Video Diffusion XT (25 frames) · fp16 · CPU offload | no reviewed results | local |
+ 96 more don't fit your machine or budget · show them
142 of 142 catalogue models listed here · sources and licences
* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory