What is LongCat-Video-Avatar 1.6?
LongCat-Video-Avatar 1.6 is Meituan's open model for audio-driven human video, released on Hugging Face on November 13, 2026. Given a speech track and either a reference photo or a text prompt, it generates one or more people speaking or singing, with mouth, face and body moving to the audio, and it can continue a video segment by segment. Meituan says 1.6 reduces strained or overly red mouth movements and the error that builds up over long generation, so clips can run past 10 minutes while keeping identity and body structure.
It sits on the 13.6-billion-parameter LongCat-Video base, frozen and shared with version 1.5, plus two LoRA adapters of 2.35 GB each that together run an 8-step sampler. The BF16 transformer is about 30 GB and a new optional INT8 version about 15 GB; an FP8 mode needs NVIDIA's Transformer Engine. The official demos run on 2 GPUs with context parallelism.
The weights and code are MIT-licensed. Comfy-Org repackaged the weights for ComfyUI on November 18, and a community node pack runs them, with core ComfyUI support still in review.
Is LongCat-Video-Avatar 1.6 worth it?
LongCat-Video-Avatar 1.6 is the strongest open option for long talking-head video, and, like version 1.5 before it, MIT-licensed. In Meituan's own preference study of 33,051 votes it ranks 2nd of six systems at 1068.1 Elo, behind OmniHuman-2.0 at 1236.7 and ahead of HeyGen at 906.4 and Kling Avatar 2.0 at 854.7.
Those numbers deserve two caveats. Meituan ran the study, and nobody has repeated it; and the gain over version 1.5, at 1051.8, is only about 16 points. The real change in 1.6 is endurance: clips past 10 minutes without the face or body drifting, which is what a lecture, a product walkthrough or a dubbed interview needs.
The catch: this is a research pipeline, not a service. The official recipe runs on 2 GPUs, the transformer alone is about 30 GB in BF16 or 15 GB in INT8, and Meituan publishes no output resolution, frame rate or VRAM figure. There is no hosted API from Meituan, and ComfyUI support depends on a repack and a community node until core support lands. Anyone using it should also think about consent: it animates any face from one photo.
Our take: OmniHuman-2.0 still looks better, but LongCat 1.6 is the one you can run on your own hardware and ship without a licence negotiation. For a training team with GPUs and hours of narration, that is a large difference.
Where it earns its place
An online-course company with 40 hours of recorded narration
Put a consistent presenter on screen for every lesson without filming
Pair one approved photo of a consenting presenter with each lesson's audio and generate continuous segments of 10 minutes or more on two in-house GPUs; cost is GPU time, with no per-minute licence fee.
A localisation studio dubbing interviews into Spanish and Japanese
Re-sync speakers' lips to the dubbed track
Feed a frame of each speaker and the dubbed audio, using the multi-person mode for two-person scenes; check the mouth region on long takes before delivery.
Before you commit
- Setup
- Clone meituan-longcat/LongCat-Video, install PyTorch, FlashAttention and the avatar requirements, download the LongCat-Video and Avatar-1.6 weights, then run the torchrun demos on 2 GPUs.
- Integrations
- Python pipeline with diffusers and transformers; Comfy-Org ComfyUI repack with a community node pack; community MLX port for Apple Silicon.
- Data and privacy
- Runs entirely on your own hardware, so photos and voice tracks stay local; because it animates any face from one photo, collect consent for the people you depict.
- Cost at scale
- No licence or per-minute fee; cost is GPU time on at least 2 GPUs for the official recipe, with INT8 or FP8 to cut memory.
- Limits
- Output resolution, frame rate and VRAM needs not published; INT8 and FP8 modes cannot be combined; first runs with compilation are slower.
What’s good
- MIT-licensed weights and code, with no regional or revenue limits
- Talking-head clips past 10 minutes without identity drift
- 2nd of six in Meituan's study, ahead of HeyGen and Kling Avatar 2.0
What’s not
- Benchmark is Meituan's own; gain over 1.5 is about 16 points
- Official demos need 2 GPUs; resolution, frame rate and VRAM not published
- No hosted API; ComfyUI support via repack and community nodes only
Best for
Teams with their own GPUs who need long, lip-synced presenter or lecture video under a permissive licence.
LongCat-Video-Avatar 1.6 alternatives
LongCat-Video-Avatar 1.6: common questions
- Is LongCat-Video-Avatar 1.6 worth it?
- LongCat-Video-Avatar 1.6 is the open talking-head model to beat, so long as you have two GPUs and no need for a hosted API. We rate it 3.8 out of 5.
- How much does LongCat-Video-Avatar 1.6 cost?
- Free weights and code under the MIT licence; no hosted API from Meituan.
- What are the downsides of LongCat-Video-Avatar 1.6?
- Benchmark is Meituan's own; gain over 1.5 is about 16 points; official demos need 2 GPUs; resolution, frame rate and VRAM not published; no hosted API; ComfyUI support via repack and community nodes only.
- Who is LongCat-Video-Avatar 1.6 best for?
- Teams with their own GPUs who need long, lip-synced presenter or lecture video under a permissive licence.
- How much does LongCat-Video-Avatar 1.6 cost for a team or in production?
- No licence or per-minute fee; cost is GPU time on at least 2 GPUs for the official recipe, with INT8 or FP8 to cut memory.
- Is LongCat-Video-Avatar 1.6 safe for company data?
- Runs entirely on your own hardware, so photos and voice tracks stay local; because it animates any face from one photo, collect consent for the people you depict.
- What are the best LongCat-Video-Avatar 1.6 alternatives?
- The closest alternatives we have reviewed are LongCat-Video, HeyGen Video 1, MiniMax H3 (Hailuo 3.0), HyperFlow H3 and LTX-2.5.
Answers checked Nov 21, 2026.