Meituan's LongCat team released LongCat-Video-Avatar 1.6 on November 15, aiming to make audio-driven synthetic video more natural and stable across extended sequences. The official Hugging Face page supplies a browsable checkpoint repository, an MIT license label and links to an open inference codebase. This is materially different from a hosted-only demonstration: researchers can inspect the component inventory and attempt their own deployment, provided they account for the foundational video model and its dependencies. Primary source ↗
The update targets synchronized lip motion, expressive faces, coordinated head and body movement and temporal consistency as generated sequences become longer. Meituan says its continuation pipeline supports video exceeding ten minutes with reduced error accumulation. A public file tree lists approximately 77.6 GB, including a base model, an optional INT8 weight-only DiT and two LoRA adapters associated with DMD and on-policy distillation. That size is a download inventory, not a measured peak GPU memory requirement.
The publisher also describes a human-preference evaluation involving six systems and 33,051 votes over fifteen pairings. Meituan reports that its new model ranked second, behind OmniHuman-2.0 and ahead of several commercial offerings and LongCat-Video-Avatar 1.5. Such rankings depend heavily on clip selection and judge instructions, and the published outcome has not been independently reproduced by Open Model Weights. It should not be represented as a general video-quality guarantee for every prompt, language or duration.
The open release nevertheless provides a meaningful verification path: developers can pin a repository revision, inspect the license and weight files, and run the published inference workflow. They must still test their own hardware footprint, pipeline reliability and output quality. Audio-driven avatars additionally require care with likeness, authorization and impersonation risks regardless of licensing. The most important fact for this registry is that new version 1.6 artifacts are available today, while performance claims remain attributed to their publisher.
What this changes for the evidence layer
Verify pinned Hugging Face file revision and MIT license text, optional INT8 DiT and two LoRA manifests, and dependency on the base LongCat-Video checkpoint. The repo contains roughly 77.6 GB and several components; do not infer runtime VRAM or consumer hardware compatibility from download volume. Human-preference ranking remains a publisher evaluation.
Meituan LongCat — official Hugging Face model card and GitHub code
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗