Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

Meituan releases LongCat-Video-Avatar 1.6

Meituan announced MIT-licensed Avatar 1.6 weights on November 15, with longer-video claims and company tests documenting remaining lip-sync errors.

Published 16 Nov 2026· 4 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

Meituan’s LongCat team announced LongCat-Video-Avatar 1.6 on November 15, 2026, releasing an audio-driven avatar video model with downloadable weights and inference code. The team says it improves speaking-face animation and consistency in videos longer than ten minutes. For university researchers and media teams, the release provides a locally runnable system to evaluate, alongside company-reported evidence that also documents persistent lip-sync errors.

What Meituan released on November 15

The dated notice in the LongCat-Video repository identifies version 1.6 as an upgrade to its audio-driven human video framework. The announcement follows Avatar 1.5, which the same release history dates to May 21, 2026. This development concerns generated video of speaking characters, a separate product from the LongCat-2.5 language-model weights CASRAI previously reported.

The announcement date differs from the upload milestone. Hugging Face’s repository metadata records creation on November 13 and a last modification that day. November 15 is the release-announcement date stated by the developers; November 13 identifies the earlier repository activity. Both fall within the current reporting week.

The model card makes the weights available under MIT terms. That makes this an open weights model that teams can download and examine. Its stated improvements concern mouth animation, appearance during extended generation, and inference efficiency.

Longer video is the central technical claim

Meituan says 1.6 reduces strained or exaggerated lip movements, excessive red coloration around the mouth and temporal instability. It also claims reduced error accumulation during continued generation, supporting videos exceeding ten minutes while maintaining identity and visual coherence. These are the developer’s claims about output quality, rather than guarantees for every recording.

The official release materials describe audio-conditioned generation with text or image inputs and video continuation. The distinction matters for research communication: the model animates a character around supplied audio. Its release does not establish that the resulting presentation is factually correct or suitable for a particular research audience.

The code documentation says 1.6 keeps the Avatar 1.5 architecture and adds DMD and OPSD adapters, with an optional INT8 base model. It also warns that FP8 and INT8 modes cannot be enabled together, and that compilation adds overhead to initial runs. Those details make the release an implementation package with configuration choices, rather than a single interchangeable checkpoint.

The published setup specifies Python 3.10, PyTorch 2.6.0 with CUDA 12.4, and FlashAttention 2.7.4.post1. These versions give university computing teams a concrete example environment to assess; they do not establish the speed or capacity of a deployment on their own infrastructure.

The company’s evaluations show gains and remaining errors

Meituan’s technical report describes an arena with 33,051 valid individual votes across all 15 pairings of six systems. Version 1.6 ranked second behind OmniHuman-2.0. A separate predecessor comparison exposes the limits of treating that rank as a complete quality assessment.

Developer-reported measure Result What it establishes
Preference against Avatar 1.5 55.6% win rate, with ties half-weighted A preference advantage within this comparison
Human-likeness on animated inputs 2.246 for 1.5; 2.114 for 1.6 The newer version did not improve every input category
Lip-sync mismatch at half-speed review 83.7% for 1.5; 82.5% for 1.6 Strict inspection still detected frequent synchronization errors

The report’s human-likeness benchmark contains 508 image–audio pairs. The authors explain that half-speed playback makes the lip-sync check unusually demanding. The figures should therefore stay attached to that review method; they do not establish a normal-speed failure rate for every ten-minute video. All these measurements come from Meituan’s evaluation, without independent replication established by the cited reporting.

Outside coverage separates availability from validation

SLOP TV News reported the November 15 release and explicitly described the preference study as Meituan’s own. Its November 16 article said it had found no independent benchmark or hands-on test of 1.6. That is the outlet’s assessment of available evidence at publication.

AI Ping’s November 15 coverage likewise distinguished the earlier weights upload from the announcement and highlighted the demanding lip-sync assessment. The two reports corroborate the release while leaving the performance evidence attributed to its developer.

What the release means for university evaluation

For researchers studying avatar generation, the useful change is access to a specific new version that can be compared with its predecessor. The published findings support separate questions about viewer preference, animated characters and synchronization. An experiment focused only on an overall rank would miss the variation documented in the report.

For university computing services, configuration is part of the evaluation record. Recording the checkpoint, adapters, quantization mode and execution environment connects the release to a reproducible AI experiment. It also allows a later comparison to distinguish model changes from implementation changes.

Research offices and media teams have a different boundary to consider. The model weight licence describes reuse of the released parameters. Separately, the model card says downstream applications have not been comprehensively evaluated and identifies accuracy, safety, fairness, privacy and content safety as considerations. The immediate development is an available research and production tool; suitability for a particular institutional use remains a separate assessment.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →