[Enterprise] Complete MiniMax H3 production qualification

[yifeif-nv]

Summary

Complete production qualification and performance checkout for MiniMax H3 through TensorRT Model Connect.

Tasks

• Confirm the applicable model license, usage terms, and artifact-distribution boundaries before model-dependent validation or publication.
• Pin the exact checkpoint, immutable revision, configuration, and source revision.
• Validate bundle generation through trtmc build.
• Validate the generated bundle through the native C++ runtime without Python or PyTorch.
• Define representative resolution, frame count, quality, precision, and runtime settings.
• Define the agreed Windows RTX and Jetson Thor hardware and software configurations.
• Benchmark build time, startup time, latency, throughput where applicable, and peak memory on each target.
• Compare output quality against the pinned reference implementation using documented metrics and visual review criteria.
• Record valid-output rate, determinism expectations, reliability, and failure behavior.
• Close identified performance, memory, quality, and runtime gaps without relaxing qualification thresholds.
• Add model-owned qualification coverage and deployment documentation.
• Document supported configurations, best-effort configurations, and known limitations.

Acceptance criteria

A versioned MiniMax H3 bundle produces validated output through the native TensorRT Model Connect runtime on the agreed Windows RTX and Jetson Thor platforms.

The exact checkpoint, configuration, quality, performance, memory use, reliability evidence, licensing boundaries, and known limitations are documented.