Meituan’s LongCat team released LongCat-Video-Avatar 1.6, an open-source framework for audio-driven human video generation, in an entry dated 2026-11-15. According to the technical report, the model ranks second among six evaluated systems in a human preference Arena with 33,051 valid votes, with a Bradley–Terry Elo score of 1068.1 (arXiv). The authors also say the refinements enable videos exceeding 10 minutes while keeping visual quality and temporal consistency (arXiv).
What the release notes say
| Date | Entry, as written by the project | Source page | Read on |
|---|---|---|---|
| 2026-11-15 | We release LongCat-Video-Avatar 1.6, an upgraded open-source framework for audio-driven human video generation. Building on LongCat-Video-Avatar 1.5, v1.6 delivers more natural lip movements and expressive facial, head, and body motion, while reducing error accumulation during long-form generation. A refined post-training pipeline combining reinforcement learning, improved classifier-free guidance for DMD-based few-step distillation, long-video rollouts, and on-policy distillation enables videos exceeding 10 minutes with improved visual quality and temporal consistency. Comprehensive inference optimizations further accelerate end-to-end generation and deployment. [code / 🤗 weights / project page ] | https://raw.githubusercontent.com/meituan-longcat/LongCat-Video/HEAD/README.md | 2026-11-18 |
| 2026-05-21 | We release LongCat-Video-Avatar-1.5, an upgraded open-source framework for audio-driven human video generation. v1.5 replaces Wav2Vec2 with Whisper-Large for more accurate lip synchronization, achieves production-ready physical rationality and temporal stability with robust long-video generation, generalizes to stylized domains (anime, animals, complex real-world conditions), supports both single-stream and multi-stream audio inputs, and accelerates inference to 8 steps via step distillation. [code / 🤗 weights / project page ] | https://raw.githubusercontent.com/meituan-longcat/LongCat-Video/HEAD/README.md | 2026-11-18 |
| 2025-12-16 | We are excited to announce the release of LongCat-Video-Avatar, a unified model that delivers expressive and highly dynamic audio-driven character animation, supporting native tasks including Audio-Text-to-Video, Audio-Text-Image-to-Video, and Video Continuation with seamless compatibility for both single-stream and multi-stream audio inputs. The release includes our Technical Report, inference code, 🤗 model weights, and project page. | https://raw.githubusercontent.com/meituan-longcat/LongCat-Video/HEAD/README.md | 2026-11-18 |
| 2025-10-25 | We’ve released LongCat-Video, a foundational video generation model. Tech report and models are available at LongCat-Video Technical Report and 🤗 Huggingface ! | https://raw.githubusercontent.com/meituan-longcat/LongCat-Video/HEAD/README.md | 2026-11-18 |
The project’s own release history, shown above, places version 1.6 as the newest of four entries. The 1.6 entry says it builds on LongCat-Video-Avatar 1.5 and delivers more natural lip movements and more expressive facial, head and body motion. It also says error accumulation during long-form generation is reduced, and that inference optimizations speed up end-to-end generation and deployment. The earlier 1.5 entry, dated 2026-05-21, described a switch from Wav2Vec2 to Whisper-Large and inference accelerated to 8 steps via step distillation. That same 1.5 entry also listed support for both single-stream and multi-stream audio inputs, and generalization to stylized domains such as anime and animals. Reading the table, 1.6 arrives about six months after 1.5 and shifts the emphasis from the audio encoder and step count toward post-training and long-form stability. The original LongCat-Video-Avatar entry, dated 2025-12-16, listed Audio-Text-to-Video, Audio-Text-Image-to-Video and Video Continuation as native tasks, and its release bundled a technical report, inference code, model weights and a project page. The base LongCat-Video entry, dated 2025-10-25, announced a foundational video generation model, so the avatar line has had three releases since that starting point.
The post-training recipe
The technical report attributes the gains to a refined post-training pipeline. Four ingredients are named: reinforcement learning, better classifier-free guidance for few-step distillation based on DMD, rollouts over long videos, and distillation performed on-policy. The report describes a train-inference mismatch in which the few-step generator is distilled on individual clips with contexts from real videos, while long rollouts feed it contexts from its own outputs, and it addresses this by training directly on such rollouts.
Arena numbers in detail
In the Arena, OmniHuman-2.0 scores 1236.7, ahead of LongCat-Video-Avatar 1.6 at 1068.1. Behind 1.6 come LongCat-Video-Avatar 1.5 at 1051.8, HeyGen at 908.1, OmniHuman-1.5 at 880.6 and Kling Avatar 2.0 at 854.7. The evaluation set consists of 508 image–speech pairs from EvalTalker. In a direct comparison, LongCat-Video-Avatar 1.6 received 55.6% of 2,535 valid votes against 44.4% for LongCat-Video-Avatar 1.5. Against OmniHuman-2.0, the report says a gap remains: OmniHuman-2.0 wins 75.3% of direct comparisons with 1.6. Our reading: the 1.6 lead over its predecessor is real but modest, at 1068.1 versus 1051.8 Elo and a 55.6% head-to-head share, and the authors’ human-likeness ratings add a caveat, since the mean for realistic inputs rises from 3.083 to 3.441 while the mean for animated inputs falls from 2.246 to 2.114, which they link to scarce long animated videos in their OPSD training set, and the distance to OmniHuman-2.0 is far larger.
Weights licence and use restrictions
The model weights are released under the MIT License, but the licence statement says it does not grant any rights to use Meituan trademarks or patents (Meituan LongCat, GitHub). The project page, however, states that generated content is for academic use only and that commercial use is not permitted (meigen-ai.github.io). Teams considering the open weights for production work should therefore read the licence and the page’s restriction on generated content together.
How we built this
Release historyDated release entries published by the project
Dated entries read by our code from the project's own release notes, news list or changelog, newest first, with the page each was read from; the entry text is the project's own wording, not a check of what it says. Each entry is shown whole. A link in an entry is shown as its text; the source page has it.
Sources
The sources this article was written from. Links open the original page.
- arXivLongCat-Video-Avatar 1.6 Technical Reportarxiv.org
- Meituan LongCatmeituan-longcat/LongCat-Video-Avatar-1.6huggingface.co
- arXivLongCat-Video-Avatar 1.6 Technical Reportarxiv.org
- meigen-ai.github.iohttps://meigen-ai.github.io/LongCat-Video-Avatar-1.6-Page: page linked by the paper as its resourcemeigen-ai.github.io
- GitHubhttps://github.com/meituan-longcat/LongCat-Video: page linked by the paper as its resourceraw.githubusercontent.com