FOUNDING OFFERSign up to unlock 35% off your first plan.

What Is LongCat-Video-Avatar 1.6? Audio-Driven Video & Production Guide

SUMMARY

LongCat-Video-Avatar 1.6 is an open-source, audio-driven human video generation model released by Meituan on November 15, 2026. Compared with version 1.5, Meituan reports reduced mouth exaggeration, less overly red lip coloration, tighter head-and-body coordination, faster inference, and reduced drift in sequences exceeding 10 minutes. While the model weights carry an MIT License on Hugging Face, the project page restricts its generated demo clips strictly to academic use. In practice, these capabilities support audio-driven digital humans and talking avatars for short product introductions or service explainers, though usable ads still depend on clear messaging, accurate product visuals, editing, and review. For independent business owners who prefer not to manage models themselves, Stillmore turns existing business photos into publish-ready videos, images, and captions; its creative team may evaluate and use models such as LongCat-Video-Avatar 1.6 when they suit a brief alongside editing and review, while you handle publishing.

Lan He

Stillmore Blog cover with a lavender gradient, the title What is LongCat Avatar 1.6, and a sculptural interpretation of the green LongCat mark; editorial artwork, not a model-generated sample.
Contents

What LongCat-Video-Avatar 1.6 Is and When It Released

LongCat-Video-Avatar 1.6 is the November 15, 2026 release in Meituan's open-source framework for generating animated human video directly from audio, text, and reference images.

The model builds on the base LongCat-Video architecture and supports English (en) and Chinese (zh) audio inputs across three native tasks. According to the LongCat-Video-Avatar 1.6 model card, those tasks cover audio-text-to-video, audio-image-text-to-video, and audio-driven video continuation.

Release notes in the official LongCat-Video GitHub repository record November 15, 2026 as the public launch date for version 1.6, following version 1.5 on May 21, 2026 and the original LongCat-Video-Avatar release on December 16, 2025. Although third-party news digests circulated earlier in mid-November, the repository changelog serves as the primary release record.

For a local dental clinic reviewing the announcement, knowing what the repository ships helps set realistic expectations. Instead of a finished ad editor, Meituan provides model code and weights designed to animate a speaking or performing figure from an audio track and visual inputs.

Why Version 1.6 Matters for Lip Naturalness and Video Length

Version 1.6 targets two recurring artifacts in synthetic talking-head video: unnatural mouth motion and visual degradation as clips run longer.

Meituan states in its model card that earlier outputs could show strained mouth movements, overly red lip coloration, and temporal instability around the mouth. On the LongCat-Video-Avatar 1.6 project page, matched-input comparisons against version 1.5 showcase calmer speech articulation alongside broader expressive categories including music, performance, speech, expression, live streaming, animation, and interaction.

Developer documentation also reports that version 1.6 mitigates autoregressive error accumulation so sequences can exceed 10 minutes while preserving visual quality and identity consistency. Still, technical duration capacity is not a marketing recommendation: a store promotion centered on one product, menu item, or service update usually communicates faster in a brief feed clip than in a 10-minute monologue.

Suppose an independent skincare brand is deciding how to introduce a seasonal moisturizer. Smoother lip sync only matters if staff specifically want a speaking presenter built from audio tracks; everyday product announcements still depend on legible labels, accurate packaging colors, and a concise hook.

Dimension LongCat-Video-Avatar 1.5 (May 21, 2026) LongCat-Video-Avatar 1.6 (Nov 15, 2026) Reported Significance
Release record Official framework update in May 2026 Official release on November 15, 2026 Latest milestone in Meituan's open-weight avatar line
Mouth and facial motion Upgraded audio encoder to Whisper-Large for lip sync Meituan reports reduced mouth exaggeration, less overly red lip coloration, and tighter head-body coordination Addresses distracting artifacts around the speaker's mouth
Long-video stability Positioned for long-video stability across domains Meituan reports reduced error accumulation for videos exceeding 10 minutes Aims to keep facial identity steady in extended clips
Inference pipeline Accelerated inference to 8 steps via step distillation Combines reinforcement learning, improved CFG for DMD distillation, long-video rollouts, and on-policy distillation Intended to speed up end-to-end generation and serving

Source: Summarized from Meituan's official GitHub README and Hugging Face model card.

How LongCat-Video-Avatar 1.6 Processes Audio and Reduces Drift

LongCat-Video-Avatar 1.6 combines audio-driven animation with a refined post-training and distillation pipeline that Meituan reports keeps long clips coherent while speeding up inference.

In autoregressive video generation, small visual errors in one frame tend to compound over time, causing faces, hands, or backgrounds to drift. According to the LongCat-Video GitHub changelog, version 1.6 counters this through reinforcement learning, improved classifier-free guidance (CFG) for Distribution Matching Distillation (DMD) few-step distillation, long-video rollouts, and on-policy distillation.

The Hugging Face model card also notes end-to-end pipeline and model execution optimizations intended to make deployment faster. Because Meituan has not published minimum GPU specs or independent throughput benchmarks in these release summaries, real-world rendering speed and setup effort depend on the operator's infrastructure.

Imagine a two-person e-commerce team weighing whether to run open-weight video tools in-house. Even with faster inference baked into the model, someone still has to configure compatible environments, prepare clean audio tracks, rerun failed generations, and check frames for hand or object distortion before publishing anything.

Licensing Rules, Demo Media Restrictions, and Production Reality

Open-weight model files under an MIT License do not make official demo clips free to reuse in ads or remove compliance responsibilities.

The LongCat-Video-Avatar 1.6 model card releases the model weights under the MIT License, noting that this grants no rights to Meituan trademarks or patents and advising downstream users to evaluate accuracy, safety, privacy, and applicable laws. Open weights let developers inspect and run the software, not repurpose sample renders for campaigns.

In fact, the ethical considerations notice on the official project page explains that parts of the sample images and audio derive from real videos solely to demonstrate research capabilities. It explicitly specifies that the generated content is for academic use only and that commercial use is not permitted.

If a boutique hotel manager browses the preview gallery, downloading a sample video to overlay a weekend room rate would violate that stated academic-only boundary. Turning an open-weight model into commercial posts requires generating assets from properly cleared source materials and checking every frame for accuracy before it goes live.

How Meituan Evaluates Version 1.6 Against Other Avatar Systems

Meituan reports that LongCat-Video-Avatar 1.6 ranked second in an internal human-preference Arena evaluation against its predecessor and four commercial systems.

According to the evaluation section on the Hugging Face model card, Meituan collected 33,051 valid individual votes across all 15 pairings among six systems. In that developer-conducted study, LongCat-Video-Avatar 1.6 posted a Bradley-Terry Elo score of 1068.1, placing behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5, and Kling Avatar 2.0.

The project page additionally presents matched-input clips comparing version 1.6 with version 1.5 and with those four commercial systems. Because the test design and clip selections come from Meituan rather than an independent third-party benchmark, they represent developer-reported results rather than proof that one system outperforms others across every business workflow.

For a neighborhood fitness studio reading those numbers, an Arena Elo score does not reveal how accurately a workflow handles class schedules, studio signage, or lighting consistency without manual editing. Treating developer benchmarks as directional context keeps production decisions grounded.

System Category in Meituan Study Reported Placement in Meituan Arena Evaluation Context
OmniHuman-2.0 Commercial system Ranked 1st Developer-reported 33,051-vote all-pairs study; exact score not listed in text
LongCat-Video-Avatar 1.6 Open-weight model (MIT weights) Ranked 2nd (Elo 1068.1) Evaluated by Meituan across 15 pairings among six systems
LongCat-Video-Avatar 1.5 Open-weight predecessor Ranked behind v1.6 Included in Meituan's matched-input and Arena comparisons
HeyGen Commercial system Ranked behind v1.6 Assessed under Meituan's selected Arena prompts and clips
OmniHuman-1.5 Commercial system Ranked behind v1.6 Included in Meituan's 15-pairing human preference study
Kling Avatar 2.0 Commercial system Ranked behind v1.6 Included in Meituan's 15-pairing human preference study

Source: Evaluation data reported by Meituan on the LongCat-Video-Avatar 1.6 Hugging Face model card and project page. Not an independent third-party benchmark.

How Audio-Driven Digital Humans Fit Real Marketing Production

Audio-driven models like LongCat-Video-Avatar 1.6 enable talking digital-human and presenter-led clips, yet turning a speaking figure into usable marketing content still depends on pairing the message with accurate product visuals, pacing, captions, editing, and review.

In practice, the capabilities showcased on the LongCat-Video-Avatar 1.6 project page—covering speech, performance, facial expression, and character interaction—center on animating a human figure from an audio track rather than building a complete business video on their own. Because talking-avatar capability and its reported benchmarks do not establish accuracy for a particular merchant's merchandise, packaging, or storefront details, any commercial clip still needs to incorporate accurate product references and be checked against what you actually sell.

Consider an illustrative 20-second introduction for a small bakery announcing a seasonal box, or a local bike shop explaining a pre-summer tune-up. A presenter segment can open with one concrete spoken takeaway—what the seasonal assortment includes or what the service check covers—while accompanying shots show close-up pastry textures, legible box labels, or the workshop bench so customers see the real item. For the final clip to communicate clearly in a social feed, the spoken hook, readable on-screen captions, and product cutaways must land in rhythm without awkward crop transitions or distorted details.

Delivering usable marketing content therefore requires editorial and post-production decisions beyond generating mouth movements: shaping the message, choosing an appropriate presenter treatment, integrating accurate product visuals, pacing the edit, writing captions, and reviewing every frame. The table below outlines how those pieces come together across common business video formats.

Video Scenario Presenter & Spoken Message Supporting Product or Space Visuals Key Production & Review Checks
Short product introduction Brief spoken hook naming one benefit, flavor, or seasonal release Close-up product shots from real photos showing packaging, color, and texture Verify label text, proportions, and surface color stay accurate across cuts
Service explanation Clear presenter summary of what the service covers and who it fits Photos of the workspace, tools, staff, or finished outcome to anchor trust Confirm on-screen captions match spoken offer terms and booking details
Store or menu update Quick update on new arrivals, weekend hours, or limited specials Storefront, display counter, or plated menu photos paired with clean pacing Check that price or date overlays stay readable long enough before the cut
Presenter-led FAQ clip Direct spoken answer to one recurring customer question Supporting detail frames illustrating the product feature being explained Inspect lip sync, hand posture, and background consistency before export

For independent business owners who would rather focus on what to promote than coordinate scripts, generation software, and video editors, Stillmore handles creative production as an advertising creative agency backed by photography and live-action, post-production and VFX, and AI production teams, as detailed in Inside Stillmore: Our Creative Production Workflow. You upload existing product, space, or team photos once and either choose a visual direction from the Stillmore Inspiration library or let the team manage ongoing selection, compositing, captioning, and pre-delivery checks while you decide when and where to post. When new models such as LongCat-Video-Avatar 1.6 emerge, Stillmore's team may evaluate and use them when they suit a brief, pairing generation with editing and creative review to pursue the best result for the piece.

Choosing Your Next Step With Audio-Driven Video Tools

Deciding how to respond to LongCat-Video-Avatar 1.6 comes down to separating technical model exploration from what your business needs to publish each week.

Engineering teams interested in testing MIT-licensed audio-driven animation can start with the repository code, task definitions, and usage considerations documented on the Hugging Face model card using cleared audio and portrait inputs.

For independent businesses focused on turning real merchandise, spaces, and services into consistent social posts for Instagram, TikTok, or YouTube Shorts, raw lip-sync duration is rarely the missing piece. Gathering sharp, well-lit existing photos of your products and storefront gives any creative workflow reliable visual proof to build on, keeping your marketing grounded in what customers actually buy rather than chasing every new model release.

Resources

Resource Direct Link What It Verifies
Meituan LongCat-Video GitHub Repository GitHub: meituan-longcat/LongCat-Video November 15, 2026 release date, version history (v1.0, v1.5, v1.6), and post-training distillation notes
Hugging Face Model Card (LongCat-Video-Avatar-1.6) Hugging Face: LongCat-Video-Avatar-1.6 Base model (LongCat-Video), MIT weight license, supported tasks, usage considerations, and Meituan's 33,051-vote Arena study
LongCat-Video-Avatar 1.6 Project Page Project Page: LongCat-Video-Avatar 1.6 Expressive scenario demos, v1.5 vs v1.6 visual comparisons, and the academic-use-only restriction on demo media
Inside Stillmore Production Workflow Inside Stillmore: Our Creative Production Workflow Photography, post-production/VFX, and AI production team roles, tool selection, and pre-delivery product checks
Stillmore Inspiration Library Stillmore Inspiration Ready-to-request creative styles built around uploaded business photos
Stillmore FAQ Stillmore FAQ Manual Inspiration requests, "Stillmore creates for me" photo thresholds, credits, delivery cadence, and owner-controlled publishing

Frequently Asked Questions

What is LongCat-Video-Avatar 1.6, and when was it released?

LongCat-Video-Avatar 1.6 is an open-source, audio-driven human video generation framework released by Meituan on November 15, 2026, according to its official LongCat-Video repository. Built on LongCat-Video, it supports English and Chinese audio-text-to-video, audio-image-text-to-video, and audio-driven video continuation with model weights released under the MIT License. Those capabilities center on generating speaking or performing human figures, whereas Stillmore works as an advertising creative agency that turns existing business photos into finished videos, images, and captions without requiring owners to install software.

What changed in LongCat-Video-Avatar 1.6 compared to version 1.5?

Meituan reports that version 1.6 reduces exaggerated mouth motion, overly red lip coloration, and mouth-area instability compared with earlier outputs, while tightening facial, head, and body coordination. It also introduces reinforcement learning, improved CFG for DMD distillation, long-video rollouts, and on-policy distillation for faster inference and reduced drift beyond 10 minutes. For businesses working with Stillmore, owners never have to follow model changelogs themselves because the creative team monitors visual trends, updates Inspiration styles, and chooses production tools behind the scenes.

Does LongCat-Video-Avatar 1.6's 10-minute capability mean business ads should be longer?

No. Meituan's note that version 1.6 can maintain stability beyond 10 minutes describes a technical milestone against frame drift rather than creative advice for small-business marketing. When introducing a product, service update, or seasonal menu item, a brief clip with clear product framing and readable captions usually communicates faster in social feeds than an extended talking segment. Through Stillmore, business owners receive concise, publish-ready videos, images, and captions built around their real merchandise and storefront.

Can businesses use LongCat-Video-Avatar 1.6 model weights commercially?

The Hugging Face model card tags LongCat-Video-Avatar 1.6 model weights under the MIT License, which permits commercial use of the weights while excluding Meituan trademarks or patents and leaving downstream users responsible for privacy, safety, and legal compliance. However, that license does not authorize reusing the project page's academic-only sample videos, and talking-avatar capability and its reported benchmarks do not establish accuracy for a particular merchant's product details, so production still has to use and check against accurate product references. Stillmore builds deliverables around your own cleared business photos rather than third-party demo assets.

Can I use the demo videos from the LongCat-Video-Avatar 1.6 project page in my ads?

No. The ethical considerations section on Meituan's project page states that parts of the sample images and audio come from real videos solely for research demonstration, and that the generated content is restricted to academic use with commercial reuse prohibited. Business owners looking for commercial-ready visual directions can browse the Stillmore Inspiration library to choose styles created around their own uploaded product, space, or team photos.

How reliable is Meituan's Arena ranking against HeyGen and OmniHuman-2.0?

Meituan reports an all-pairs human preference study of 33,051 valid votes across 15 pairings and six systems, placing LongCat-Video-Avatar 1.6 second with a Bradley-Terry Elo score of 1068.1 behind OmniHuman-2.0 and ahead of v1.5, HeyGen, OmniHuman-1.5, and Kling Avatar 2.0. Because Meituan conducted the study internally rather than through an independent lab, those scores do not measure how accurately a workflow handles product labels, pacing, or brand consistency. Stillmore evaluates tools internally and checks deliverables against your photos before delivery.

Why isn't generating lip-synced speech enough on its own for a marketing video?

Generating synchronized mouth movement addresses only one layer of a video, while effective product introductions and service explainers also depend on a focused message, natural presenter framing, accurate visuals of the item or space, edit pacing, captions, and visual review. Because talking-avatar capability and its reported benchmarks do not establish accuracy for a particular merchant's product details, Stillmore's photography and live-action, post-production and VFX, and AI production teams use accurate product references during compositing and editing and check deliverables before delivery.

Might Stillmore use a model like LongCat-Video-Avatar 1.6 for client videos?

Yes, conditionally: Stillmore's team may evaluate and use models such as LongCat-Video-Avatar 1.6 when they suit a specific brief, combining generation with post-production editing and creative review to pursue the best result for the piece. At the same time, Stillmore does not tie deliverables to one named model or offer a standalone avatar or voice-cloning service; clients receive finished videos, images, and matching captions crafted from their existing business photos.

What photos do I need to get started with Stillmore?

Clear everyday photos taken on your phone or stored on your computer can supply the starting material, provided key details such as labels, product shapes, and surface colors are easy to see. You upload those photos once to your Library so they can be reused across future promotions. To enable "Stillmore creates for me", upload at least 3 photos on Try it or 15 photos on Growth (monthly or annual); below that threshold, you can still pick an Inspiration and submit requests manually.

How do Stillmore's delivery cadence, credits, and automatic creation work?

With Growth in automatic mode, your first video arrives within 5 days once you have at least 15 suitable photos in Library, "Stillmore creates for me" enabled, completed payment, and sufficient Credits, with a normal schedule of one video every 5 days, as detailed in Stillmore's FAQ. Requests you submit yourself—including videos and images—are also delivered within 5 days once payment, required materials, and your request are complete. Monthly Growth Credits cover up to 6 videos total (3,000 Credits per video, 500 Credits per image) across both modes; late-month uploads may mean remaining work cannot all be delivered before month-end.

Who publishes the finished videos and owns the social schedule?

The business owner keeps full control over publishing. Stillmore handles photo selection, creative direction, production, caption writing, and pre-delivery review, then notifies you when finished videos or images are ready in your account. You choose when and where to post them—such as Instagram, TikTok, or YouTube Shorts—because Stillmore does not auto-post to your social channels or guarantee views or sales.

About the author

Lan He

I'm Lan. Years of writing promotional and editorial work in the US, mostly for companies with big audiences and bigger expectations — where a single article might reach hundreds of thousands of readers, and where the good ones kept earning traffic long after the campaign ended. I learned what makes a page survive that: one clear audience, one real question, images that show the actual work, and a structure that keeps people reading. Now I write for small business owners who do everything themselves, answering the questions they actually ask: what's worth posting, which photos are worth taking, and how to tell one story across images, video, and words. I've seen good content carry a small company further than its ad budget ever did.

More about Lan He & all articles ↗

Give them a reasonto stop by.

Send us your phone photos. Get posts ready to share.

See what we can make
Stillmore fashion creative look
All stories ↗