whatmodel.sh / video / Wan 2.1 14B
Wan 2.1 14B
↗ share
+ contribute
Wan 2.1 14B is an open-weights video model from Alibaba, released February 2025. Its published size is 14B parameters. Licence: apache-2.0.
$ how it compares
The video models ranked either side of Wan 2.1 14B, out of 139 with a score, then more from Alibaba.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 78 | Wan2.1-I2V-14B-720P | 48±4 | comparison only | ||
| 79 |  Cosmos-Predict2-2B-Video2World | 48±8 | comparison only | ||
| 80 | Wan 2.1 14B | 47±2 | api $0.24/sec +local | ||
| 81 | Wan 2.2 TI2V 5B | 47±2 | api price unknown +local | ||
| 82 |  Kandinsky 5.0 T2V Lite | 47±2 | local | ||
| more from Alibaba | |||||
| 2 | Wan 3.0 Prime | 80±7 | api $0.14/sec | ||
| 4 | Wan 3.0 | 79±2 | api $0.10/sec | ||
| 9 | HappyHorse 1.1 | 76±3 | api $0.14/sec | ||
$ what it's good at
text to video▆▄▃▄▆▅█▃▂▂▂▄ #36 of 68 · 88% of its score · 11 tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VideoPhy-2 (human, joint SA & PC) | 32.6% ✎ | ▄···▄·▄·█··▄ beats 5/5 | 28% |
| Cosmos HUE human evaluation, text-to-video (Cosmos 3 report) | 84.0% ✎ | █▄▄··▄▄▄▄▄·▄ beats 3/9 | 19% |
| SkyReels-Bench T2V (human rating) | 62.4% ✎ | ▄····▄···█·▄ beats 4/4 | 15% |
| T2VPhysBench (human physical consistency) | 42.0% ✎ | █··█·█·····█ beats 3/3 | 11% |
| VBench (T2V total) lab | 86.2% ✎ | █···▄▄▄····▄ beats 5/5 | 8.4% |
| VBench T2V total (long-prompt version, SkyReels-V2 runs) | 83.7% ✎ | █···█···█··█ beats 3/3 | 7.4% |
| VBench T2V — Subject consistency (AMD sample sets) | 93.7% ✎ | no other model yet | — |
| VBench T2V — Background consistency (AMD sample sets) | 94.8% ✎ | no other model yet | — |
3 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VBench T2V — Motion smoothness (AMD sample sets) | 98.9% ✎ | no other model yet | — |
| VBench T2V — Aesthetic quality (AMD sample sets) | 62.7% ✎ | no other model yet | — |
| VBench T2V — Imaging quality (AMD sample sets) | 65.6% ✎ | no other model yet | — |
other tests1 not in a task · 9.5% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Wan-Bench (Wan2.1 release) weighted score lab | 72.4% ✎ | █·······█··█ beats 2/2 | 9.5% |
$ where to run it
Through an API
api · Replicate · $0.24/sec · replicate.com ↗ · price sourceUSD per second of output video, 720p: Replicate official model wavespeedai/wan-2.1-t2v-720p ('Accelerated inference for Wan 2.1 14B text to video') is billed $0.24 per second of output video (billing table on the model page); the 480p model wavespeedai/wan-2.1-t2v-480p is $0.07 per output second. Product: Replicate model API (WaveSpeedAI accelerated serving). API model: wavespeedai/wan-2.1-t2v-720p. Settings: text-to-video, 720p. Price basis: published output-based list rate (no hardware-time charge). Currency: USD. Not an Alibaba direct tariff. Replicate official model page, observed 2026-09-25.
Other ways to run it
version
- Run it as published · needs ~20.2 GB · from github.com/leejet
sd-cli -M vid_gen --diffusion-model models/wan2.1-t2v-14b-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 6.0 --sampling-method euler -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 --flow-shift 3.0
Weights and documentation: huggingface.co/Wan-AI ↗
$ how the score was worked out · VideoPhy / VideoPhy-2 officia… 28%, SkyReels-V2: Infinite-length… 23%, Cosmos 3: Omnimodal World Mod… 19% and 3 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
| source | results | share of the score |
|---|---|---|
 VideoPhy / VideoPhy-2 official repository leaderboards | 1 | 28% |
 SkyReels-V2: Infinite-length Film Generative Model | 2 | 23% |
| Cosmos 3: Omnimodal World Models for Physical AI (arXiv:2606.02800v4) | 1 | 19% |
| T2VPhysBench: A First-Principles Benchmark for Physical Consistency in Text-to-Video Generation | 1 | 11% |
 Wan-AI/Wan2.1-T2V-14B model card · lab-reported | 1 | 9% |
| Alibaba Wan technical report (arXiv:2503.20314v2), VBench table (own models) · lab-reported | 1 | 8% |
| prior · pull toward the middle | 2% | |
| weighted average | = 47.4 |
not counted (5)
| Breaking the Accuracy-Speed Barrier: How MXFP4/6 Quantization Revolutionizes Image and Video Generation | VBench T2V — Subject consistency (AMD sample sets) [correct] | 93.7% | fewer than three models have this test yet, or this source ran it for this model alone |
| Breaking the Accuracy-Speed Barrier: How MXFP4/6 Quantization Revolutionizes Image and Video Generation | VBench T2V — Background consistency (AMD sample sets) [correct] | 94.8% | fewer than three models have this test yet, or this source ran it for this model alone |
| Breaking the Accuracy-Speed Barrier: How MXFP4/6 Quantization Revolutionizes Image and Video Generation | VBench T2V — Motion smoothness (AMD sample sets) [correct] | 98.9% | fewer than three models have this test yet, or this source ran it for this model alone |
| Breaking the Accuracy-Speed Barrier: How MXFP4/6 Quantization Revolutionizes Image and Video Generation | VBench T2V — Aesthetic quality (AMD sample sets) [correct] | 62.7% | fewer than three models have this test yet, or this source ran it for this model alone |
| Breaking the Accuracy-Speed Barrier: How MXFP4/6 Quantization Revolutionizes Image and Video Generation | VBench T2V — Imaging quality (AMD sample sets) [correct] | 65.6% | fewer than three models have this test yet, or this source ran it for this model alone |
$ published speed measurements · 2
Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.
| NVIDIA H800/H100 | 1837.9 s/video | Wan reference implementation · T2V-14B; 720P; 1 GPU; --offload_model True; no prompt extension; 50 sampling steps PNG time column manually visually verified. Combined GPU labels preserved; no claim that the run was specifically the H100 rather than H800. CPU model offload is explicit in settings; not misclassified as no offload. | source ↗ | |
| NVIDIA A800/A100 | 3342.6 s/video | Wan reference implementation · T2V-14B; 720P; 1 GPU; --offload_model True; no prompt extension; 50 sampling steps PNG time column manually visually verified. Combined GPU labels preserved; no claim that the run was specifically the H100 rather than H800. CPU model offload is explicit in settings; not misclassified as no offload. | source ↗ |
$ improve this page
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