Meituan releases LongCat-Video-Avatar-1.6, open-weight avatar model for 10-minute-plus videos under MIT license
Meituan has released LongCat-Video-Avatar-1.6, an open-source audio-driven human video generation model with weights under the MIT License. The company claims it sustains videos longer than 10 minutes and ranks second among six systems in a human preference study of 33,051 valid votes.
Meituan's LongCat team has released LongCat-Video-Avatar-1.6, an open-source model for audio-driven human video generation, with weights published on Hugging Face under the MIT License. The company says it can generate videos longer than 10 minutes while preserving visual quality, identity consistency and temporal coherence.
What changed from 1.5
LongCat-Video-Avatar-1.6 builds on LongCat-Video-Avatar 1.5 and uses the same base architecture. According to Meituan, it improves expressive naturalness, lip-audio synchronization, coordinated facial and body motion, and visual consistency in long-form generation. The company lists three headline changes:
- Lip animation: reduces exaggerated or strained mouth movements, overly red lip coloration, and temporal instability in the mouth region.
- Long-video stability: mitigates error accumulation during autoregressive video continuation.
- Inference speed: end-to-end pipeline and model execution optimizations for faster deployment and serving.
A technical report accompanies the release. The model supports single- and multi-character audio-driven generation. Documented modes are speech-text-to-video, speech-image-to-video, and video continuation for single-person animation, plus speech-image-to-video and video continuation for multi-person animation.
Human evaluation
Meituan compared 1.6 against its predecessor 1.5 and four commercial systems: OmniHuman-2.0, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0. The all-pairs preference study covers 15 model pairings and 33,051 valid individual votes.
According to Meituan, LongCat-Video-Avatar 1.6 ranks second of six systems with a Bradley-Terry Elo score of 1068.1. It sits behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0. The results come from the company's own evaluation.
Checkpoint and inference details
The checkpoint package includes two task-specific LoRA adapters: a DMD LoRA (loraA/lora.safetensors) and an OPSD LoRA (loraB/lora.safetensors). The --mix_steps=N option applies a blend of DMD and OPSD (weighted by mix_alpha) for the first N denoising steps, then OPSD only for the remaining 8 - N steps. Defaults are --mix_steps=4 and --mix_alpha=1.0, so the first four steps use DMD only and the last four use OPSD only. Meituan says raising either value gives DMD more influence early on, which may improve structural stability and reduce visual distortions.
Other deployment notes:
- An optional weight-only INT8 DiT (
base_model_int8) is enabled with--use_int8. - FP8 (
--use_fp8) requires Transformer Engine with PyTorch and CUDA support. FP8 and INT8 cannot be enabled together. For quality-sensitive runs, the guidance is to reduce the number of layers using FP8. - FlashAttention-2 is the default. FlashAttention 3 is recommended on Hopper GPUs (H20, H100, H200, H800), where Meituan says it can give a significant speedup over FA2. xformers is also supported via config.
- With
--compile_ditand--compile_vae, the first few runs are slower because of graph tracing and compilation. New input shapes may trigger recompilation. - On more than two GPUs, setting
LONGCAT_A2A_WORKSPACE=1enables an optimized All-to-All workspace to cut context-parallel communication overhead.
The reference setup uses Python 3.10, torch 2.6.0+cu124 and flash_attn 2.7.4.post1. Users need both the foundational LongCat-Video weights and the LongCat-Video-Avatar-1.6 weights.
License
Model weights and repository contributions are under the MIT License. The license does not grant rights to Meituan trademarks or patents. Context window and API pricing are not applicable, since the release is distributed as downloadable weights.
What this means
The release gives builders an avatar video model with permissive licensing and documented knobs for speed and memory: LoRA mixing, INT8 and FP8 quantization, FlashAttention 3 and multi-GPU context parallelism. The ranking is Meituan's own study, so teams considering it against OmniHuman-2.0 or other commercial systems should test on their own audio and identity inputs. The 10-minute-plus claim targets error accumulation in autoregressive continuation, which is the main constraint in long-form avatar generation. Independent long-form results will show whether it holds in practice.
Related Articles
Shanghai AI Laboratory releases Sage S2.1 Flash, a 321B-parameter scientific multimodal model under MIT license
Shanghai Artificial Intelligence Laboratory has published Sage S2.1 Flash, a 321B-parameter scientific multimodal model with MIT-licensed weights and an FP8 variant. According to the lab, it placed first on four of 15 charted benchmarks and second on seven, including 90.3 on MLE-bench Lite.
Huawei Cloud releases CloudRobo-WM1.2 world action model, claims first place on RoboTwin 2.0 and LIBERO-Pro 2.0
Huawei Cloud has released CloudRobo-WM1.2, an embodied world action model, and says it will be open source. The company claims first place on the RoboTwin 2.0 and LIBERO-Pro 2.0 robot manipulation benchmarks, with scores of 88.98% and 81.7%.
Meituan releases LongCat-2.5: 1.6T-parameter multimodal MoE with 48B active per token
Meituan's LongCat team has published LongCat-2.5, a native multimodal mixture-of-experts model with 1.6 trillion total parameters and approximately 48 billion activated per token. It targets long-horizon agentic tasks such as codebase development and full-project builds, and ships with W4A4 expert quantization for domestic AI accelerators.
MiniMax releases MiniMax-H3-Image-PE, a prompt-expansion model fine-tuned from Qwen3.8-27B in FP4
MiniMax has released MiniMax-H3-Image-PE on Hugging Face, a prompt-expansion model for MiniMax-H3-Image. The weights are fine-tuned from Qwen3.8-27B in FP4 and turn a short request into a detailed image description plus an aspect-ratio recommendation.
Comments
Loading...