Wan 2.1
Video Available Full comparison ↗ wan-video/wan-2-1 · by Wan Video (Alibaba)
· Diffusion transformer (DiT)
Pricing — 1 offering(s)
About this offering: Wan 2.1 image-to-video on Replicate, served through WaveSpeed AI's accelerated inference at 480p and 720p, billed per second of output video. Wan 2.1 text-to-video is also on Replicate but priced on its own model pages. The newer Wan 2.5 is available on fal.ai.
Image-to-Video 480p
- $0.09 / second Current 2026-06-13 → present
Image-to-Video 720p
- $0.25 / second Current 2026-06-13 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Wan 2.1 fits into a cost-aware routing setup
See how →About Wan 2.1
Wan 2.1 is the model that established Alibaba's Wan (Tongyi) series in the open-weight video-generation community. Alibaba open-sourced it in February 2025 as a family of four checkpoints — text-to-video at 14B and 1.3B parameters, and image-to-video at 14B in 480p and 720p — with a first-and-last-frame variant following in April. Architecturally it is a flow-matching diffusion transformer paired with Wan-VAE, a 3D causal autoencoder that encodes and decodes 1080p video of arbitrary length while preserving temporal continuity, and a T5 text encoder for multilingual prompts. The Wan technical report (arXiv 2503.20314) describes training on a curated corpus of billions of images and videos; the smaller 1.3B text-to-video model was deliberately sized to run in about 8 GB of VRAM on a consumer GPU.
All four checkpoints ship under the Apache 2.0 licence, weights included, so the model can be self-hosted or reached through third-party hosts — the Replicate offering in this registry is a WaveSpeedAI-accelerated image-to-video wrapper around those weights. On VBench, Alibaba reported a total score of 86.22 for the 14B model (86.67 quality, 84.44 semantic), and Wan 2.1 topped the public VBench leaderboard at release; Modelglass records those as the vendor's launch figures rather than an independent re-run and rates the rest of this entry's capability profile as a qualitative read. As an Alibaba model it is developed and hosted primarily in China, which can carry data-residency or procurement constraints for some buyers.
Wan 2.1 has since been succeeded twice, and the distinction matters when choosing a target to build against. Wan 2.2 (July 2025) is the direct open-weight successor — a mixture-of-experts redesign, still Apache 2.0 — and is the upgrade path where a host offers it. Wan 2.5 (September 2025) added synchronised native audio, 1080p output and ten-second clips but is API-only: Alibaba did not release its weights, breaking the open-weight line. Route to Wan 2.1 when you specifically need open weights you can run yourself, or the low per-second cost of the accelerated 480p image-to-video path; for a hosted text-to-video job with no self-hosting requirement, a newer Wan release is usually the better pick.
Capability profile
How Wan 2.1 rates across core capability dimensions, with the task-level evidence behind each rating.
VBench quality score 86.67 for the 14B model at release; leads the dynamic-degree and multi-object sub-dimensions.
VBench semantic score 84.44 for the 14B model at release.
No audio.
WaveSpeed-accelerated hosting on Replicate provides fast I2V inference.
Replicate offering is I2V only; T2V available via self-hosting.
T2V and I2V only; no editing of existing clips via Replicate or fal.ai.
Benchmarks
| Benchmark | Score | Config | Source |
|---|---|---|---|
| vbench | 86.22 | Figures as reported by Alibaba at the February 2025 open-source release, when Wan 2.1 led the public VBench leaderboard. Verify against the current leaderboard as entries are updated. | leaderboard ↗ |
Operator guidance
Wan 2.1 on Replicate is optimised for I2V at $0.09/s (480p) or $0.25/s (720p). For T2V, prefer Wan 2.5 via fal.ai ($0.05/s) or self-hosting. For I2V at production volume, Wan 2.1 480p on Replicate is very cost effective.
Use cases
- Image-to-video generation at cost-effective rates via Replicate
- Self-hosted T2V and I2V pipelines (open-weight)
- High-volume I2V processing (480p for speed, 720p for quality)
Limitations
- Replicate offering is I2V only via the WaveSpeedAI accelerated models
- No synchronised native audio (a video-to-audio task exists in the open-weight release but is not exposed by the hosted offerings tracked here)
- Superseded as an open-weight target by Wan 2.2 (July 2025, MoE, still Apache-2.0); Wan 2.5 (September 2025) is a further generation but is API-only, with no public weights
- VBench figures are the vendor's launch numbers, not an independent re-run; the rest of the capability profile is qualitative