LongCat-Video-Avatar 1.6 generates avatars past 10 minutes
The MIT-licensed release keeps version 1.5's architecture, adds two LoRA adapters, and ranks second of six in Meituan's own preference study, behind ByteDance's OmniHuman-2.0.

Key takeaways
- Meituan's LongCat-Video-Avatar 1.6 generates videos longer than 10 minutes while holding visual quality, identity and timing steady, the model card and technical report say.
- The checkpoint reuses the 1.5 architecture and adds two LoRA adapters plus an optional INT8 version of the base model, under the MIT license.
- In Meituan's own all-pairs study of 33,051 votes, LongCat-Video-Avatar 1.6 scores a Bradley-Terry Elo of 1068.1, second of six behind ByteDance's OmniHuman-2.0.
- No independent benchmark or hands-on test of LongCat-Video-Avatar 1.6 had been published as of November 16, 2026.
Meituan's LongCat team shipped LongCat-Video-Avatar 1.6 on November 15, an open-weight model that animates people speaking from a supplied audio track.
The release is announced in the LongCat-Video GitHub repository and documented on the model's Hugging Face card and in a technical report. The card says version 1.6 produces videos exceeding 10 minutes while preserving visual quality, identity consistency and temporal coherence.
Three things changed from version 1.5. Mouth movement is calmer, with fewer strained shapes and less of the over-saturated red lip tint that appeared as a clip ran long. The generator dropped from 50 denoising steps to eight through a DMD2-based distillation. The report traces both fixes to a post-training pipeline that pairs reinforcement learning with on-policy self-distillation over long-video rollouts.
LongCat-Video-Avatar is the team's family of open audio-driven avatar models, built on LongCat-Video, its 13.6B-parameter video generator. Version 1.5 arrived on May 21, 2026, adding a Whisper-Large audio encoder.
Meituan's own human-likeness study, on 508 image-and-speech pairs, lifts 1.6 above 1.5 on a five-point scale, from 3.006 to 3.322 overall. The report's expert review cuts the aggregate consistency issue rate from 19.9% to 11.6%. On animated inputs the score falls, from 2.246 to 2.114, which the report attributes to thin coverage of long animated videos in its training set.
The report places 1.6 in a six-model preference arena drawn from 33,051 valid votes across all 15 pairings. It scores a Bradley-Terry Elo of 1068.1, second behind ByteDance's closed OmniHuman-2.0 at 1236.7 and ahead of version 1.5 (1051.8), HeyGen (908.1), OmniHuman-1.5 (880.6) and Kling Avatar 2.0 (854.7). Meituan ran and analyzed that study itself. HeyGen bills by subscription, from $29 a month on its own pricing page, read November 16, 2026, where the MIT weights carry no per-second fee.
The checkpoint keeps 1.5's architecture and adds two LoRA adapters, a DMD LoRA and an OPSD LoRA, plus an optional INT8 base model. On Hugging Face the base weights ship as six safetensors shards totaling about 31.7 GB; the INT8 set is four shards at about 15.9 GB. The card and report state no minimum VRAM, and every example command runs on two GPUs.
No independent benchmark or hands-on test of 1.6 had been published as of November 16. The hands-on figures in circulation are for version 1.5: Tight Studio measured a peak of 46,827 MiB on one H200, and VisionStory kept roughly 40,500 MiB resident on an A800. The open MIT license makes 1.6 the option to watch for talking-head work, though the setup still assumes datacenter GPUs.
LongCat-Video-Avatar 1.6 is on Hugging Face and a ModelScope mirror under the MIT license, with the inference code in the LongCat-Video repository; running it needs Python 3.10, PyTorch 2.6.0 with CUDA 12.4 and flash-attn 2.7.4.post1.
Sources
- huggingface.co - model card: features, license, arena, usage note
- github.com - release announcement, model list, requirements
- github.com - Elo table, distillation, human-likeness and issue rates
- meigen-ai.github.io - official project page
- huggingface.co - weight file sizes
- github.com - 1.6 merge commit, dated November 15, 2026 02:57 UTC
- heygen.com - HeyGen's own plan prices, read November 16, 2026
- tight.studio - independent H200 test of LongCat-Video-Avatar 1.5
- visionstory.ai - independent A800 test of LongCat-Video-Avatar 1.5
- theaiengineering.com - release tracker confirming the November 15, 2026 date and open weights
- reddit.com - community reaction