Wan2.7-260612 is a specialized video generation model variant within Alibaba’s Wan 2.7 model family, released on June 12, 2026. Developed by Alibaba's Tongyi Lab, the model is built on a Mixture-of-Experts (MoE) Diffusion Transformer (DiT) architecture, with a total model capacity of 27B MoE parameters. Wan2.7-260612 sits at the upper tier of the AI video generation market, achieving competitive performance on public leaderboards for its realistic motion dynamics, temporal consistency, and high-fidelity output of up to 1080p resolution and up to 15 seconds in duration.
The model’s most notable architectural feature is its Thinking Mode. Rather than treating user prompts as basic triggers, the model uses an internal reasoning step to plan the video's layout, lighting, composition, and physical movement prior to generation. This reasoning phase ensures that the final video adheres closely to complicated instruction briefs, resulting in fewer generation attempts and more deliberate cinematic outcomes.
Wan2.7-260612 supports a highly flexible multimodal pipeline that includes text-to-video (T2V), image-to-video (I2V), reference-to-video (R2V), and instruction-based video editing. It provides advanced First & Last Frame Control, allowing creators to feed the model a start and end image and automatically generate the motion sequence between them. For consistent character generation, the R2V feature lets creators supply reference images or videos to preserve a character’s face, body, and even vocal timbre across multiple distinct shots. Furthermore, its precise audio-to-video sync enables automatically coordinated lip-sync, character motion, and ambient rhythm matching when an audio track is supplied.
Prompting Guide & Tips
- Leverage the Thinking Mode: Write highly detailed briefs describing the scene's subject, lighting, environment, and camera work. The model processes these details sequentially to construct the scene with high intent.
- Control Sequences: Use specific formatting to dictate motion paths, such as naming a starting layout and an ending perspective.
- Avoid Hype Terms: Instead of using subjective quality terms, explicitly detail high-contrast lighting, film grain, or lens details to direct the model's visual rendering.
- Apply Negative Prompts: Use negative prompts to specify what to avoid, such as low-quality artifacts, watermarks, text overlay distortion, or motion blur, keeping the output clean and cinematic.