AlibabaVideo model
Wan 2.2 A14B
In image to video, Wan 2.2 A14B is in place 19 of 21, with text only in place 22. A 5 s clip costs about $0.40.
1,170Arena ratinghigher is better
1,132Arena ratinghigher is better
$0.40$0.08 per secondlower is better
5slongest single cliphigher is better
27out of 100-1 vs. last month
Summary
At $0.08 per second of video it is cheap; a typical video model costs $0.12.
What is distinctive about Wan 2.2 A14B is its open-weight MoE branch with 27 billion parameters and 14 billion active per computation step. It handles text-to-video and image-to-video at 720p for up to 5 seconds, explicitly without native audio.
Compared with Wan 2.1, Alibaba cites more training data, stronger cinematic control, and more complex motion. Unlike hosted Wan 2.6, this branch can be run locally.
Wan 2.2 A14B benchmarks
Image to video arena
How people rate generated videos in blind comparisons, Elo points · higher is better
20 of 21 models
as of September 22, 2026 · Source: LMArena
Text to video arena
How people rate generated videos in blind comparisons, Elo points · higher is better
20 of 23 models
as of September 22, 2026 · Source: LMArena
Show all 97 individual results from 12 sources
Page 1 of 9
Published benchmarks
LMArena Text to Video
Elo ratings summarize pairwise comparisons within this leaderboard. They are not percentages. The scale does not express absolute quality or linear gaps between models.
higher is betterAs of November 5, 2026Scale: 0 Elo - 1,516 Elo
Methodology and context
- Metric
- elo-rating
- Scale
- No upper bound is documented. The chart spans the recorded values and zero.
- Participants
- 23
- Sample
- 10,458
LMArena Image to Video
Elo ratings summarize pairwise comparisons within this leaderboard. They are not percentages. The scale does not express absolute quality or linear gaps between models.
higher is betterAs of November 5, 2026Scale: 0 Elo - 1,495 Elo
Methodology and context
- Metric
- elo-rating
- Scale
- No upper bound is documented. The chart spans the recorded values and zero.
- Participants
- 21
- Sample
- 27,054
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
VideoPhy-2 detailed subset joint success (Show measurement, test conditions, and source)
- Source value
- 47
- Score
- 47
- Metric
- Joint Semantic-Adherence-and-Physical-Commonsense success: both ratings are at least 4 on a 1-5 scale
- Unit
- %
- Benchmark version
- arXiv:2608.29904v1; VideoPhy-2 detailed subset, Table 1
- Category
- video-generation:physical-commonsense
- Direction
- Higher is better
- Sample
- 400
- Source type
- independent-evaluator
- Evaluator
- Off-Manifold Refinement: Guiding Video Generators with a Frozen World Model
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Base Wan2.2-T2V-A14B text-to-video sampler, explicitly named in the paper abstract and bridged to the assigned profile by the official Wan-AI T2V-A14B model card. Table 1 labels the evaluated base row Wan2.2. The default primary evaluation uses a fixed 400-prompt VideoPhy-2 detailed subset with one generation per prompt; joint success requires both Semantic Adherence and Physical Commonsense ratings to be at least 4 on a 1-5 scale. No OMR-enhanced or image-to-video score is transferred.
- Context
- The source gives no confidence interval or other statistical uncertainty interval, so none is reported. The evaluation identifies the base T2V-A14B sampler but does not publish an immutable weight-file digest.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 19.6
- Score
- 19.6
- Metric
- Acc_QA (question-answer accuracy)
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 37 (shared 80-case set)
- Direction
- Higher is better
- Sample
- 80
- Lower confidence bound
- 14.8
- Upper confidence bound
- 24.9
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 17.5
- Score
- 17.5
- Metric
- Score_PR (Acc_QA^0.8 × s_dyn^0.2)
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 37 (shared 80-case set)
- Direction
- Higher is better
- Sample
- 80
- Lower confidence bound
- 12.5
- Upper confidence bound
- 23.1
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 30
- Score
- 30
- Metric
- S(v), VLM-judged video quality weighted 0.4 reasoning, 0.3 temporal consistency, 0.3 aesthetics
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 37 (shared 80-case set)
- Direction
- Higher is better
- Sample
- 80
- Lower confidence bound
- 25.9
- Upper confidence bound
- 34.6
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 21.6
- Score
- 21.6
- Metric
- Acc_QA, Diff. (implicit, high-level intent) prompt regime
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Table 3 (Diff. (implicit, high-level intent))
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. QA accuracy under the source's separate prompt regimes: Diff. gives only high-level intent; Easy adds explicit transition guidance. Table 3 does not state regime-specific N or CI.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 35.2
- Score
- 35.2
- Metric
- Acc_QA, Easy (hinted, explicit transition guidance) prompt regime
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Table 3 (Easy (hinted, explicit transition guidance))
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. QA accuracy under the source's separate prompt regimes: Diff. gives only high-level intent; Easy adds explicit transition guidance. Table 3 does not state regime-specific N or CI.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 22.9
- Score
- 22.9
- Metric
- Score_PR, World-Knowledge
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 38 (shared evaluation set)
- Direction
- Higher is better
- Sample
- 21
- Confidence interval
- 14.5
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 14.5
- Score
- 14.5
- Metric
- Score_PR, Human-Centric
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 38 (shared evaluation set)
- Direction
- Higher is better
- Sample
- 9
- Confidence interval
- 12.6
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 16.4
- Score
- 16.4
- Metric
- Score_PR, Logic-Reasoning
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 38 (shared evaluation set)
- Direction
- Higher is better
- Sample
- 27
- Confidence interval
- 7.8
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
WorldReasonBench (Show measurement, test conditions, and source)
- Source value
- 15
- Score
- 15
- Metric
- Score_PR, Information-Based
- Unit
- %
- Benchmark version
- arXiv:2605.10434v1, Appendix Table 38 (shared evaluation set)
- Direction
- Higher is better
- Sample
- 23
- Confidence interval
- 7.8
- Source type
- independent-evaluator
- Evaluator
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. WorldReasonBench open-source model evaluation; exact source row Wan2.2-14B maps to the assigned wan-2-2-a14b profile. Appendix P says open-source models followed their official inference recipes on H100 nodes, generating 5-second 480p-720p clips from image-plus-text inputs, one generation per case under each of two prompt regimes. The paper gives no immutable checkpoint hash or serving route.
Recorded measurement details
LMArena Text to Video: LMArena Text to Video (Show measurement, test conditions, and source)
- Source value
- 1132
- Score
- 1132
- Metric
- elo-rating
- Unit
- Elo
- Task
- lmarena-text-to-video
- Task label
- LMArena Text to Video
- Direction
- Higher is better
- Source rank
- 42
- Sample
- 10458
- Lower confidence bound
- 1117
- Upper confidence bound
- 1147
- Retrieved at
- 2026-11-05
LMArena Image to Video: LMArena Image to Video (Show measurement, test conditions, and source)
- Source value
- 1170
- Score
- 1170
- Metric
- elo-rating
- Unit
- Elo
- Task
- lmarena-image-to-video
- Task label
- LMArena Image to Video
- Direction
- Higher is better
- Source rank
- 43
- Sample
- 27054
- Lower confidence bound
- 1160
- Upper confidence bound
- 1180
- Retrieved at
- 2026-11-05
Page 1 of 9
Video sample from our test
This video exists as a documented run for the exact model version. Playback is muted.
Motion and physical consistency
This sample tests Anatomy, object consistency, camera tracking, rain, and reflections.
Prompt and settings
Generate exactly one continuous five-second 16:9 cinematic video. A fictional bicycle courier in a mustard-yellow raincoat rides a red city bicycle from left to right across a rain-wet Hamburg street at blue hour for the entire clip. The camera tracks her smoothly at waist height. Both hands remain visible on the handlebars, the bicycle geometry stays physically correct, and a transparent umbrella fixed to the bicycle moves naturally in the wind. Wet-street reflections and raindrops follow realistic physics. No cuts, no slow motion, no text, no logos, no subtitles, and no audio.
Generated through fal.ai with model fal-ai/wan/v2.2-a14b/text-to-video. Archived settings were 16:9, one request, first output, requested 5 s, seed 958910179, request {"prompt":"Generate exactly one continuous five-second 16:9 cinematic video. A fictional bicycle courier in a mustard-yellow raincoat rides a red city bicycle from left to right across a rain-wet Hamburg street at blue hour for the entire clip. The camera tracks her smoothly at waist height. Both hands remain visible on the handlebars, the bicycle geometry stays physically correct, and a transparent umbrella fixed to the bicycle moves naturally in the wind. Wet-street reflections and raindrops follow realistic physics. No cuts, no slow motion, no text, no logos, no subtitles, and no audio.","num_frames":81,"frames_per_second":16,"adjust_fps_for_interpolation":false,"num_interpolated_frames":0,"interpolator_model":"none","resolution":"720p","aspect_ratio":"16:9","enable_safety_checker":true,"enable_output_safety_checker":true,"enable_prompt_expansion":false}, configured price estimate $0.40.
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
20 of 98 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 5, 2026
Wan 2.2 A14BPixVerse V5.6Grok Imagine Video 1.5Kling 2.6 ProWan 2.6
Wan 2.2 A14B compared
Compare with another model
Predecessor and successor
The most popular comparisons
- 1Gemini Omni Flash vs. Wan 2.2 A14B
- 2FLUX.3 Video vs. Wan 2.2 A14B
- 3Dreamina Seedance 2.5 720p vs. Wan 2.2 A14B
- 4Dreamina Seedance 2.0 720p vs. Wan 2.2 A14B
- 5MiniMax H3 vs. Wan 2.2 A14B
- 6Sora 2 Pro vs. Wan 2.2 A14B
- 7Sora 2 vs. Wan 2.2 A14B
- 8Veo 3.1 vs. Wan 2.2 A14B
- 9Veo 3.1 Fast vs. Wan 2.2 A14B
Show all 43 comparisons
- 10Grok Imagine Video 1.5 vs. Wan 2.2 A14B
- 11Grok Imagine Video vs. Wan 2.2 A14B
- 12Wan 2.7 vs. Wan 2.2 A14B
- 13Runway Gen-4.5 vs. Wan 2.2 A14B
- 14Kling 3.0 Pro vs. Wan 2.2 A14B
- 15Kling 2.6 Pro vs. Wan 2.2 A14B
- 16Vidu Q3 Pro vs. Wan 2.2 A14B
- 17PixVerse V5.6 vs. Wan 2.2 A14B
- 18Hailuo 2.3 vs. Wan 2.2 A14B
- 19LTX-2 19B vs. Wan 2.2 A14B
- 20Wan 2.2 A14B vs. HunyuanVideo 1.5
- 21Wan 2.2 A14B vs. Pika 2.2
- 22Wan 2.2 A14B vs. Gemini Omni 1.1 Flash
- 23Wan 2.2 A14B vs. Muse Video
- 24Wan 2.2 A14B vs. Wan 3.0
- 25Wan 2.2 A14B vs. Wan 3.0 Prime
- 26Wan 2.2 A14B vs. HappyHorse 1.0
- 27Wan 2.2 A14B vs. HappyHorse 1.1
- 28Wan 2.2 A14B vs. PixVerse V6
- 29Wan 2.2 A14B vs. PixVerse C1
- 30Wan 2.2 A14B vs. Seedance 2.0 Fast
- 31Wan 2.2 A14B vs. Seedance 2.0 Mini
- 32Wan 2.2 A14B vs. Vidu Q3 Turbo
- 33Wan 2.2 A14B vs. Vidu Q2
- 34Wan 2.2 A14B vs. Vidu Q1
- 35Wan 2.2 A14B vs. Kling 4.0
- 36Wan 2.2 A14B vs. Kling 3.0 Standard API
- 37Wan 2.2 A14B vs. Kling Video O1
- 38Wan 2.2 A14B vs. Wan 2.5 Preview
- 39Wan 2.2 A14B vs. Wan 2.2 TI2V 5B
- 40Wan 2.2 A14B vs. HunyuanVideo 1.0 T2V
- 41Wan 2.2 A14B vs. Ray 3.2
- 42Wan 2.2 A14B vs. Grok Imagine Video 1.5 Lite
- 43Wan 2.2 A14B vs. LTX-2.5
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
Prices
No Alibaba subscription includes this model. Through other providers from ≈ €21.20 a month (Recraft Pro, with limited credits). 2 plans in Germany.
| Plan | Monthly price | Note |
|---|---|---|
| Recraft Pro | ≈ €21.20 | with limited credit |
| Recraft Teams | ≈ €23.33 | with limited credit, for teams, per seat |
Frequently asked questions about Wan 2.2 A14B
How good is Wan 2.2 A14B?
Rank 22 of 23 in the text-to-video arena rating. With a start image: rank 19 of 21.
LMArena, as of Sep 22, 2026
How much does a video with Wan 2.2 A14B cost?
About $0.40 for a 5 s clip, $0.08 per second of video. That is cheap.
How long can a video with Wan 2.2 A14B be?
Up to 5 seconds in one go. Resolution: 720p.
According to Alibaba
- Price
- $0.08 per second$0.40 per 5 s clip
- Longest clip
- 5s
- Resolution
- 720p
- Audio
- no
- Open weights
- yes
Where can I use Wan 2.2 A14B?
Not in a Alibaba subscription. Including other providers that makes 2 plans in Germany.
Subscription comparison, prices in Germany incl. VAT
At Alibaba
No Alibaba subscription includes this model.
Through other providers
2 plans from 1 provider, from ≈ €21.20 a month (Recraft Pro), with limited credits.
Sources and data date
Show all 28 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
| Evaluator | Test series | Composite indices | Recorded active measurements |
|---|---|---|---|
| LMArena | 0 | 0 | 1 |
Without vendor claims and historical or disputed results; those stay labelled in the single values. 1 measurements are not yet assigned to a test series.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 16, 2026. 22 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | wiro-video-models-exact-row Note The inspected Wiro leaderboard has no exact model-version route. Nearby variants were not transferred. |
| Not found in the inspected sources | arena-video-edit-exact-row Note The inspected Video Edit table has no exact model-version row. T2V and I2V values were not reused. |
| Not found in the inspected sources | vabench-exact-model-row Note VABench has no exact model row. Sora 2 was not transferred to Sora 2 Pro or other models. |
| Not found in the inspected sources | artificial-analysis-t2v-exact-row Note No exact model version or resolution was found in the inspected Artificial Analysis table. |
| Not found in the inspected sources | artificial-analysis-i2v-exact-row Note No exact model version or resolution was found in the inspected Artificial Analysis table. |
| Unresolved | vbench-current-public-leaderboard Note The current VBench leaderboard is embedded in a dynamic Hugging Face iframe and was not retrievable; current scores remain unresolved. |
| Not found in the inspected sources | vidbenchmark-exact-row Note No exact model id is present in the 2026-08-04 dataset; similarly named variants were not transferred. |
| Not published | immutable-checkpoint-and-serving-route Original wording The source identifies Wan2.2-14B and describes the open-source inference setup, but does not give an immutable weight hash or serving route. |
| Not published | table-3-regime-sample-count-and-intervals Original wording Table 3 gives Diff. and Easy QA accuracy but no per-regime sample count or confidence interval; none is inferred from other tables. |
| Not found in the inspected sources | VBench2 score Note The official VBench2 leaderboard lists no exact Wan2.2 T2V-A14B row. Its Wan2.2 I2V-A14B entry is a separate image-to-video model and is not mapped. |
| Additional source | Alibaba (retrieved October 17, 2026) · Wan2.2 T2V A14B model card · Editorial description reviewed October 2, 2026 |
| Additional source | Video pricing |
| Additional source | fal.ai |
| Additional source | LMArena (retrieved November 5, 2026)12 |
| Additional source | MLCommons MLPerf Inference v6.0 (retrieved October 16, 2026) |
| Additional source | RBench, Rethinking Video Generation Model for the Embodied World (retrieved October 17, 2026; October 24, 2026) |
| Additional source | Off-Manifold Refinement: Guiding Video Generators with a Frozen World Model (retrieved October 17, 2026) |
| Additional source | WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (arXiv:2605.10434v1) (retrieved October 18, 2026) |
| Additional source | Xu et al., Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE (arXiv:2609.38140v1) (retrieved October 20, 2026) |
| Additional source | VGIF-Score paper, independent evaluator (retrieved October 22, 2026) |
| Additional source | AWS Neuron Documentation, Wan2.2-T2V-A14B Model Card (retrieved October 23, 2026) |
| Additional source | RoboTrustBench (retrieved October 23, 2026) |
| Additional source | Arena Video Text-to-Video (retrieved September 21, 2026; October 16, 2026; October 28, 2026) |
| Additional source | VBench official leaderboard (retrieved February 18, 2026; October 28, 2026) |
| Additional source | Model metadata (retrieved October 16, 2026)1234 |
| Additional source | Model metadata (retrieved October 28, 2026) |
| Additional source | Alibaba Cloud Model Studio video generation and editing · Editorial description reviewed October 2, 2026 |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |
| Additional source | Alibaba Cloud Community: "Alibaba Unveils Wan2.6 Series Enabling Everyone to Star in Videos" (2026-09-19) |