LongCat-Video-Avatar 1.6 is an open-weights model from Meituan's LongCat team that animates a still portrait from an audio clip to produce talking-head video. It is an upgrade to version 1.5 with more natural lip movement, steadier facial and body motion, and less drift during long generation, which the team says allows clips longer than 10 minutes. Code is on GitHub and the weights are released under the MIT license. In a human preference study of 33,051 votes across six systems it ranked second, behind OmniHuman-2.0.
A workable dense video generator from Meituan. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | VRAM |
|---|---|---|
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770)ACEMAGIC | SS | 0.5 GB |
| Acer Veriton GN100 AI MiniAcer | SS | 0.5 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 0.5 GB |
| AMD Instinct MI300XAMD | SS | 0.5 GB |
| AMD Instinct MI325XAMD | SS | 0.5 GB |
Cheapest current cloud rentals with at least 1 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.05 |
NVIDIA Tesla V100 16GBVast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · Spot · 16 GB VRAM | $0.09 |
NVIDIA GeForce RTX 5070Vast.ai · Spot · 12 GB VRAM | $0.09 |
NVIDIA Tesla V100 16GBVast.ai · On-Demand · 16 GB VRAM | $0.09 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.
LongCat-Video-Avatar 1.6 is an open-weights human video synthesis model developed by Meituan's LongCat team. Designed to animate a still portrait image using an input audio clip, the model generates synchronized, identity-preserving talking-head video with expressive facial dynamics and coherent body motion. While commercial services like HeyGen and Kling Avatar dominate proprietary digital human pipelines, LongCat-Video-Avatar 1.6 provides an accessible, locally deployable alternative released under the permissive MIT license.
Built on top of the dense LongCat-Video base model, LongCat-Video-Avatar 1.6 features an undisclosed parameter count. The release addresses the primary failure mode of automated portrait animation: progressive visual degradation over long temporal sequences. In version 1.5, extended generations suffered from unnatural mouth shapes, strained lip openings, and color saturation drift, often causing lips to turn unnaturally red over time. Version 1.6 introduces architectural and training refinements that suppress error accumulation across autoregressive continuation steps, enabling continuous video generation exceeding 10 minutes without identity divergence or visual artifacts.
In an extensive Bradley-Terry human preference arena comprising 33,051 valid votes across six systems, LongCat-Video-Avatar 1.6 achieved an Elo rating of 1068.1, ranking second overall. It surpassed commercial systems including HeyGen, Kling Avatar 2.0, and OmniHuman-1.5, finishing behind only ByteDance's OmniHuman-2.0. For engineers deploying local avatar generation pipelines, it represents one of the strongest open-weights foundations available.
LongCat-Video-Avatar 1.6 operates as a dense architecture with undisclosed parameters, fine-tuned specifically for audio-driven visual generation from the core 13.6B LongCat-Video foundation model. Unlike Mixture of Experts (MoE) designs that route inputs dynamically through sparse sub-networks, dense visual synthesis architectures activate all weights for every forward step, ensuring maximum representation capacity and deterministic VRAM consumption across generation stages.
The core pipeline couples an audio encoder with a spatiotemporal diffusion transformer backbone. Audio features extracted from speech or music clips are injected via cross-attention layers, mapping phonetic and acoustic cadences directly to facial landmark deformations and mouth geometries. The underlying video generation engine employs a coarse-to-fine generation strategy along both the spatial and temporal axes, paired with block-sparse spatial attention to accelerate high-resolution synthesis without blowing out memory budgets.
To solve the drifting problem inherent in autoregressive video generation, Meituan restructured the video continuation mechanism. While traditional models pass reconstructed frame latents sequentially into future context windows (causing small artifacts to compound), LongCat-Video-Avatar 1.6 enforces tight identity conditioning anchors derived from the original source portrait alongside temporal stabilization losses. The effective context length is not explicitly quantified as a standard token window, but the model supports sequence continuation capable of sustaining continuous 10-minute generations at 720p resolution and 30 frames per second without losing facial geometry or wardrobe details.
LongCat-Video-Avatar 1.6 handles both single-modal and multimodal driving inputs:
Deploying video diffusion systems locally differs from running quantized text models. While developers searching for a local AI model undisclosed parameters 2026 often default to lightweight runtimes, LongCat-Video-Avatar 1.6 relies on a PyTorch and Hugging Face Diffusers pipeline rather than an LLM loader. You cannot simply load this pipeline into Ollama as the quickest way to get started for text completion; rather, it requires Python-based execution using CUDA-accelerated inference scripts.
Because LongCat-Video-Avatar 1.6 is derived from a 13.6B dense base model with heavy visual attention latents, LongCat-Video-Avatar 1.6 VRAM requirements are demanding for local workstations:
For practitioners wondering how to run undisclosed model on consumer GPU setups, keeping generations at lower initial resolutions and relying on PyTorch's torch.compile, FlashAttention-2, and sequential CPU offloading is mandatory on 24 GB cards.
When seeking the best quantization for LongCat-Video-Avatar 1.6, practitioners typically look for FP8 or INT8 weight representations. Standard 4-bit integer weights like Q4_K_M (common in GGUF-based LLMs) are not standard for diffusion transformer backbones due to visual artifacting in facial regions. Running the model in native BF16 or FP8 scaled via bitsandbytes or native PyTorch FP8 delivers the best balance of fidelity and memory reduction.
LongCat-Video-Avatar 1.6 performance is measured in frames per second (FPS) and end-to-end generation latency rather than LongCat-Video-Avatar 1.6 tokens per second. On an NVIDIA RTX 4090, generating a 5-second 720p 30fps clip requires several minutes of compute depending on diffusion sampling steps. On an enterprise H100 accelerator, generation approaches near real-time processing speeds, making it practical for batch production workloads.
Understanding where LongCat-Video-Avatar 1.6 fits requires comparing it to contemporary open and closed alternatives in the avatar synthesis ecosystem.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Meituan model we track.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.