What shipped

LongCat-Video-Avatar 1.6 is an audio-driven character video model for single- or multi-character generation, with video continuation across segments. Version 1.6 states three targets: more natural lip movement with less temporal instability in the mouth region, coordinated facial and body motion, and identity that holds as a clip is extended segment by segment.

The engineering is incremental and readable on the card. Same base architecture as 1.5. Two task-specific LoRA adapters, mixed across the first denoising steps under a --mix_steps setting, with the stated aim of structural stability early in generation. An optional weight-only INT8 DiT for deployment. None of that is exotic. The claim that matters is what it is supposed to buy: runs past ten minutes.

The ten-minute claim is a claim

Meituan says 1.6 mitigates the error accumulation that autoregressive continuation produces, enabling videos exceeding ten minutes while preserving visual quality, identity consistency and temporal coherence. What the model card does not give is a measurement a reader can check: no drift metric over time, no published comparison of minute one against minute ten, and no speed figure behind the separate claim of faster end-to-end inference. The card's gallery shows selected outputs past ten minutes. Selected outputs demonstrate that the runs exist; they do not establish how reliably identity holds across them.

There is also no hosted quick start on the listing itself. At the time of checking, Hugging Face showed no Inference Provider for the model, and the card's example commands use two GPUs, without the documentation establishing a minimum configuration. The practical position for most readers is simple: the weights are downloadable today, and the sources reviewed provide no independent evaluation of the headline capability.

Second place, in Meituan's own arena

The ranking needs the same reading. Meituan reports 33,051 valid votes across all 15 pairings of a six-system field that includes HeyGen and Kling Avatar 2.0. That is a vote count, not necessarily a count of independent raters or representative tasks. The study was run and reported by the model's maker, the pairing methodology and clip selection are Meituan's, and the sources reviewed provide no reproduction of it by an outside group.

Treat the Elo of 1068.1 as a statement about how 1.6 performed in Meituan's test, on Meituan's chosen clips, against Meituan's chosen field. It is a reason to test the model. It is not a settled ordering of the avatar market.

The license is the solid part

The least hedged fact in the release is the license. The weights are MIT: commercial use, modification and redistribution are permitted under its terms. The license permits evaluating and adapting the weights; deployment practicality is a separate question. Two cautions ride with it. MIT on the weights does not clear the inputs: a face or voice you do not have rights to stays uncleared. And Meituan's demonstration page restricts its own generated demo content to academic, noncommercial use, a separate term readers should not blur into the model license.

If you build talking-head, teaching or narration video, this is worth a local test, run at the length you actually ship, checked at the joins. If you need a hosted avatar today, the listing offers no Inference Provider route yet.

Sources

  1. [1] Meituan LongCat, LongCat-Video-Avatar 1.6 model card, Hugging FaceRead source
  2. [2] AIGC News, "Meituan LongCat open-sources LongCat-Video-Avatar-1.6," November 15, 2026Read source
  3. [3] Sysseder, "LongCat Video Avatar 1.6 targets the weak spot in long talking videos," November 16, 2026Read source