Alibaba has officially exposed its Wan 3.0 Prime text-to-video model on Replicate, offering a public endpoint for generating clips up to 30 seconds in resolution options ranging from 480p to 1080p. While the API schema supports both text-only generation and image-conditioned animation, the release is shrouded in technical ambiguity. The most glaring issue for developers is the complete absence of documented frame rates, parameter counts, or inference time guarantees. If you are looking for a drop-in replacement for existing video workflows, be warned: the available materials provide zero benchmarks for the Prime endpoint specifically.

The Documentation Gap

A critical distinction in the source material is that the supplied GitHub README describes Wan 2.1, not Wan 3.0 Prime. This means any hardware requirements listedβ€”such as the 8.19 GB VRAM usage or the four-minute generation time on an RTX 4090 for a 5-second clipβ€”are strictly for the older Wan 2.1 T2V-1.3B variant. Applying these figures to Wan 3.0 Prime is a category error. The schema for the new model exposes inputs like duration, resolution, seed, and enable_prompt_expansion, but it fails to define the exact pixel dimensions for the advertised resolutions or the specific enum values for aspect ratios. For a model marketed as 'Prime,' this lack of transparency is concerning.

Input Constraints and Image Conditioning

The endpoint accepts an optional starting image (JPG, PNG, BMP, or WebP) capped at 10 MB, which overrides the aspect_ratio setting in favor of the input image's dimensions. This allows for animating stills, such as adding camera moves to product shots. However, the schema does not support video input, multiple reference images, audio input, or manual frame-rate controls. The enable_prompt_expansion flag defaults to true, which the documentation notes improves short prompts but increases latency. Without a stated speed guarantee, this default setting could unpredictably slow down batch processing pipelines.

Competitive Landscape and Licensing Risks

When compared to sibling models like wan-2.7-t2v (which supports synchronized audio) or wan-pro (which focuses on 30 fps short clips), wan-3-prime differentiates itself primarily through its 30-second duration ceiling and optional image conditioning. Yet, the source explicitly states that no controlled quality, speed, or cost comparisons exist between these endpoints. Furthermore, the licensing situation is murky; the metadata points to the Wan 2.1 repository license, but there is no confirmation that this governs the Wan 3.0 Prime weights or the hosted service. Commercial teams must verify terms independently, as no safety policy or bias evaluation for this specific endpoint is included in the release notes.

Key Takeaways

  • Wan 3.0 Prime supports text-to-video and image-to-video generation with clips up to 30 seconds at 480p, 720p, or 1080p, but lacks documented frame rates or parameter counts.
  • Hardware specs from the accompanying GitHub README apply to Wan 2.1, not Wan 3.0 Prime, meaning developers cannot rely on existing VRAM or latency benchmarks for the new model.
  • The licensing status is unclear, as metadata references the Wan 2.1 license without confirming it applies to Wan 3.0 Prime weights or the hosted Replicate service.

The Bottom Line

Wan 3.0 Prime is a promising but under-documented tool. Until Alibaba releases actual benchmarks for the Prime endpoint, treat it as an experimental API rather than a production-ready solution.