Introduction

Talking-head video just got a new open option. Meituan's LongCat team has released LongCat-Video-Avatar 1.6, the next version of its open framework for audio-driven human video generation. You give it a voice track plus a text prompt or a reference image, and it animates one or more people speaking. The weights are on Hugging Face under the MIT license (posted November 13, 2026), and the 1.6 inference code was merged into the main branch of the LongCat-Video GitHub repo on November 15, 2026 (02:57 UTC). You can run it today on your own GPUs.

What changed from 1.5

The model card lists three main upgrades:

  • Calmer, more natural lips. Fewer exaggerated or strained mouth shapes, less of the overly red lip tint, and less jitter around the mouth.
  • Long videos that hold together. The team says it reduced the error build-up in autoregressive continuation, so videos can run past 10 minutes while keeping identity, quality and timing consistent.
  • Faster end-to-end inference. The pipeline and model execution were optimized for quicker deployment and serving.

1.6 keeps the same base architecture as 1.5. It ships with two task LoRAs and an optional quantized DiT:

  • loraA is a DMD LoRA and loraB is an OPSD LoRA, blended across an 8-step distilled schedule.
  • base_model_int8 is an optional weight-only INT8 DiT you can turn on with --use_int8. FP8 through Transformer Engine is also supported, but don't use it together with INT8.

How it scores

Meituan ran an all-pairs human preference arena with 33,051 valid votes. LongCat-Video-Avatar 1.6 scored a Bradley–Terry Elo of 1068.1, which put it second of six. It came in behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0. These are the vendor's own numbers, so treat them as a promising signal rather than a final verdict.

What it can do

The demo scripts cover these modes:

  • Speech + text to video: a talking person made from a prompt alone
  • Speech + image to video: animate your own reference portrait
  • Video continuation: chain segments into long-form clips
  • Multi-person: several characters, each driven by its own audio

The checkpoint also includes a Whisper-large-v3 audio encoder and a vocal separator, so noisy or music-backed voice tracks can be cleaned up first.

How to try it

The official path is the LongCat-Video repo with PyTorch 2.6, CUDA 12.4 and FlashAttention. You need the base LongCat-Video weights plus the 1.6 avatar checkpoint:

git clone --single-branch --branch main https://github.com/meituan-longcat/LongCat-Video
cd LongCat-Video
pip install -r requirements.txt -r requirements_avatar.txt
huggingface-cli download meituan-longcat/LongCat-Video --local-dir ./weights/LongCat-Video
huggingface-cli download meituan-longcat/LongCat-Video-Avatar-1.6 --local-dir ./weights/LongCat-Video-Avatar-1.6

Then run the single-person speech + image demo. The official examples use two GPUs with context parallelism:

torchrun --nproc_per_node=2 run_demo_avatar_single_audio_to_video.py --context_parallel_size=2 --checkpoint_dir=./weights/LongCat-Video-Avatar-1.6 --stage_1=ai2v --input_json=assets/avatar/single_example_1.json --use_distill --lora_path loraB/lora.safetensors --model_type avatar-v1.6 --mix_steps=4 --mix_alpha=1.0

For long-form output, add --num_segments=5 --ref_img_index=10 --mask_frame_range=0 to chain continuation segments. If you see structural wobble, the card suggests raising --mix_steps or --mix_alpha so the DMD LoRA has more influence in the early steps.

This is a heavy, research-grade setup. The examples assume multi-GPU and Hopper cards benefit from FlashAttention 3. ComfyUI users had community wrappers for 1.5, but I couldn't confirm 1.6 support yet, so keep an eye on those repos.

Original Source

Conclusion

Open avatar models used to give up after about a minute of talking. LongCat-Video-Avatar 1.6 goes for the marathon: 10+ minute takes, calmer lips, multi-person scenes, and an MIT license you can actually build on. If you have GPUs to spare, give it a podcast clip and see how long your digital host keeps its cool.

—Aurelia ♡