whatmodel.sh / video / Wan 2.2 I2V A14B

Wan 2.2 I2V A14B

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Wan 2.2 I2V A14B is an open-weights video model from Alibaba, released July 2025. Its published size is 27B parameters. Licence: apache-2.0.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

27B total across the two MoE experts (high-noise + low-noise), 14B active per step; diffusion model only (T5 encoder and VAE separate). Matches the catalogue convention for wan-2-2-a14b (params_b 27, active_b 14); active_b 14 if wanted.

55.8±1.8#34 of 93 video modelsfrom 9 benchmark results across 8 sources

$ how it compares

The video models ranked either side of Wan 2.2 I2V A14B, out of 139 with a score, then more from Alibaba.

#modelscorewhereRAMreleased
40Kling 2.5 Turbo Pro Kuaishou57±2api $0.070/sec2025-09
41Runway Aleph Runway56±5comparison only2025-07
42Wan 2.2 I2V A14B Alibaba56±2api $0.080/sec +local~31GB2025-07
43Seedance 1.0 Pro ByteDance Seed56±2api $0.052/sec2025-06
44Kling O1 Pro Kuaishou (Kling AI)56±2api $0.11/sec
more from Alibaba
2Wan 3.0 Prime Alibaba80±7api $0.14/sec2026-08
4Wan 3.0 Alibaba79±2api $0.10/sec2026-08
9HappyHorse 1.1 Alibaba76±3api $0.14/sec2026-06

see all video models · find the ones that fit your machine

$ what it's good at

image to video▁▂·▂▅█▅▄▅▆▃▄ #32 of 50 · 86% of its score · 7 tests

Beats Seedance 1.5 Pro here, though they rank higher overall. Trails Cosmos3-Nano, Hailuo 2.3, Wan 2.6 Flash here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
WorldReasonBench human Elo (image+text-to-video world-state prediction)953.0 ✎▃▃··█··▃▃▃·▃ beats 4/8avg 875▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 39%sharp
CRONOS counterfactual rigid-body physics (I2V success rate)20.0% ✎█·····█··█·█ beats 3/3very hard · avg 12%▇▇▇▇▇▇▇ 17%sharp
PAI-Bench-G human preference arena (Overall ELO)1082.6 ✎▄······██··█ beats 5/6avg 1002▇▇▇▇▇ 12%typical
Cosmos HUE human evaluation, image-to-video (Cosmos 3 report)88.4% ✎▄·█··▄··▄▄▄█ beats 5/8easy · avg 84%▇▇▇ 7.4%typical
VGI-Bench visual reasoning (I2V)21.6% ✎▄▄···▄▄█·█·▄ beats 1/8hard · avg 20%▇▇ 5.8%typical
PAI-Bench-G physical-AI video generation (Overall)82.3% ✎▃······▅·█▅█ beats 10/10easy · avg 81%▇ 3.3%blunt
Physics-IQ Verified (I2V, best-practice prompts)32.2% ✎██···█··█·██ beats 3/5hard · avg 25%▇ 2.3%blunt · at its level
other tests2 not in a task · 13% of its score
testresultvs the others in its rundifficultycountswhy
MindWorldBench intention accuracy (I2V mental-state reasoning)28.7% ✎▅·▃▅·█·▃···▃ beats 6/9hard · avg 21%▇▇▇▇ 9.0%at its level
PhyGround physics human evaluation (Overall, TI2V)65.6% ✎█▄▄··▄···▄·█ beats 7/7medium · avg 57%▇ 3.5%blunt · at its level

$ where to run it

Through an API

api · fal.ai · $0.080/sec · fal.ai ↗ · price sourceUSD per video second at 720p on fal.ai. Product: fal Wan 2.2 A14B image-to-video endpoint. API model: fal-ai/wan/v2.2-a14b/image-to-video. Settings: resolution=720p. fal counts video seconds at 16 frames per second. 580p $0.06, 480p $0.04 per video second. Currency: USD. fal.ai model page, observed 2026-09-25.

Other ways to run it

version

  1. Run it as published · needs ~31.2 GB · from github.com/leejet
    sd-cli -M vid_gen --diffusion-model models/Wan2.2-I2V-A14B-LowNoise-Q4_K_M.gguf --high-noise-diffusion-model models/Wan2.2-I2V-A14B-HighNoise-Q4_K_M.gguf --vae models/wan_2.1_vae.safetensors --t5xxl models/umt5-xxl-encoder-Q8_0.gguf -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 -i input.png --flow-shift 3.0

Weights and documentation: huggingface.co/Wan-AI ↗

$ how the score was worked out · WorldReasonBench: Human-Align… 39%, CRONOS: Benchmarking Counterf… 17%, PAI-Bench: A Comprehensive Be… 15% and 5 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv v1)1▇▇▇▇▇▇▇▇▇▇▇▇ 39%CC-BY-4.0
CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models (arXiv:2605.23699v1), Table 11▇▇▇▇▇ 17%CC BY 4.0
PAI-Bench: A Comprehensive Benchmark For Physical AI (arXiv:2512.01989v1)2▇▇▇▇▇ 15%CC-BY-4.0
MindWorldBench: Evaluating Mental-State-to-Behavior Reasoning in Image-to-Video Generation (arXiv:2609.39147v1)1▇▇▇ 9%CC-BY-4.0
Cosmos 3: Omnimodal World Models for Physical AI (arXiv:2606.02800v4)1▇▇ 7%CC-BY-4.0
VGI-Bench: Probing Visual Intelligence in Video Generation Models1▇▇ 6%CC-BY-4.0
PhyGround: Benchmarking Physical Reasoning in Generative World Models (arXiv:2605.10806v1)1▇ 3%CC-BY-4.0
Physics-IQ Verified1▇ 2%CC-BY-4.0
prior · pull toward the middle▇ 1%
weighted average= 55.8

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json