Published: 25 August 2026
Alibaba released its Wan3.0 video model broadly on 24 August 2026 at $0.20 per output second for 1080p, which comes to $6.00 for a full 30-second clip. That doubles the maximum length of its predecessor Wan2.5. On the same day, Alibaba Cloud’s PAI team published three open-weights SearchQwen search agents on Hugging Face.
What can Wan3.0 do?
Wan3.0 generates clips of up to 30 seconds at resolutions up to 1080p, and it takes text, images, video and audio in a single prompt. One prompt accepts up to ten images, five videos and five audio clips. Web pages and documents also work as input, including PDFs, spreadsheets and PowerPoint files, so a slide deck can go straight to video without an intermediate script.
The model recommends a clip length based on the prompt and ships an extension tool for lengthening videos that already exist. Alibaba’s stated focus is consistency, meaning characters, props and spatial layout that stay stable across the full 30 seconds, which is where most video models start to drift. Alibaba has published no benchmark or evaluation behind that claim.
Wan3.0 pricing, access and weights
There are two tiers.
- Standard costs $0.05 per output second at 480p, $0.10 at 720p, and $0.20 at 1080p, currently with a 30 percent discount applied.
- The faster Prime tier costs $0.068, $0.14 and $0.28 for the same three resolutions.
A 30-second 1080p clip therefore lands at $6.00 on Standard and $8.40 on Prime, and a 30-second 480p clip at $1.50.
Access runs through the wan.video site, Alibaba Cloud Model Studio and the Qwen Cloud API. There are no downloadable weights for Wan3.0, and the open Wan line still stops at Wan 2.2 under Apache 2.0.
Alibaba Cloud PAI also shipped three open SearchQwen models
On 24 August, Alibaba Cloud’s PAI team published SearchQwen3-8B, SearchQwen2.5-7B and SearchQwen2.5-3B on Hugging Face, alongside SynSearch-Data, a training set of 5,000 verified multi-hop trajectories covering 4,182 audited tasks. These are compact search agents that emit structured search and browse tool calls, so they need an external search backend before they do anything useful. All three ship real safetensors weights.
SearchQwen3-8B scores 35.42 on Alibaba’s deep search average against 24.50 for the untuned Qwen3-8B, and 50.31 overall against 40.31. Context is 40,960 tokens. Every comparison is against the matching base Qwen model and nothing else, using Alibaba’s own LLM-judge harness. SearchQwen3-8B and SearchQwen2.5-7B are Apache 2.0. The 3B carries the non-commercial Qwen Research licence, so only two of the three are usable commercially.
What this means
Wan3.0 is worth testing if you produce short-form video at volume and the 15-second ceiling forced you to stitch clips together. At $6.00 per 30-second 1080p clip the arithmetic works for a marketing team and stops working at real scale. Document input is the unusual part, since no competing model takes a slide deck directly. Treat the consistency claim as unproven.
SearchQwen is the more interesting release for engineering teams, and the more oversold one. An 8B model running a deep research loop on your own hardware under Apache 2.0 has real value where a hosted API is not permitted. But benchmarked only against its own base model, with no paper, the scores prove the fine-tuning worked, not that the model competes.
For more information and the full details visit the official Wan 3.0 page on the Alibaba Cloud Model Studio.
