Feature: Add Wan 3.0 video model support

This was generated by AI during triage.

Request type

Future model support consideration

Current status

Wan 3.0 has been released as a hosted product/API, but its model weights have
not been open-sourced. TensorRT-Model-Connect should not begin implementation
based on unofficial or unverifiable artifacts. Reconsider adding Wan 3.0 only
after Alibaba/Wan-AI publishes official downloadable weights and the material
needed to reproduce the reference behavior.

Problem and use case

Alibaba released Wan 3.0 as its latest video-generation model after a public
preview that began on August 6, 2026. TensorRT-Model-Connect currently lists
Wan 2.1 T2V and Wan 2.2 TI2V support, but it has no Wan 3.0 workload or model
family coverage.

Wan 3.0 is currently available through Alibaba Cloud Model Studio under the
wan3.0-video API. At the time this issue was filed, the official Wan GitHub
organization and Wan-AI model hubs did not provide open-source Wan 3.0 model
weights, a model card, or a reference implementation. Third-party repositories
that only wrap the hosted API or claim unofficial weights are not suitable
model sources.

Proposed outcome

After official Wan 3.0 weights are open-sourced, evaluate whether the model is
a suitable addition to TensorRT-Model-Connect. If it is, add a minimal,
representative path using the official artifacts:

• Pin an official model ID, revision, and license.
• Determine whether Wan 3.0 is architecture-compatible with an existing Wan
  family or requires a new model-owned family and runtime strategy.
• Build a TensorRT-Model-Connect bundle through the public builder interface.
• Run the bundle through the normal user-facing runtime path.
• Add focused validation against an official reference and update the public
  supported-model table.

Acceptance criteria

• The selected checkpoint is published by Alibaba/Wan-AI, downloadable
  without private credentials, and pinned to an immutable revision.
• The issue or implementation records the architecture/config differences
  from the existing Wan 2.1 and Wan 2.2 families.
• A representative Wan 3.0 configuration builds through the public
  trtmc interface and produces a valid bundle.
• The generated bundle runs through the normal runtime path without
  benchmark-only hooks.
• Focused correctness validation passes against the official reference
  with meaningful, behavior-aligned criteria.
• The supported-model documentation includes the new Wan 3.0 workload.
• No test threshold, passing criterion, or validator is weakened to make
  the model pass.

Out of scope

• Treating the Alibaba Cloud hosted API as a TensorRT backend.
• Using unofficial, leaked, or unverifiable model weights.
• Supporting every Wan 3.0 mode, resolution, and optimization in the first
  contribution.

Upstream references

• Official release page: https://modelstudio.console.alibabacloud.com/model-releases/wan3.0-video
• Official Model Studio overview: https://modelstudio.alibabacloud.com/intl/blog/wan3-ai-video-generation-model/
• Official Wan GitHub organization: https://github.com/Wan-Video

Submission checks

• I searched open and closed issues and found no duplicate request.
• I removed secrets, private/internal evidence, personal paths, and
  restricted artifacts.